image_pack 0.2.0 → 0.2.2

data/ext/image_pack/image_pack.c

@@ -16,6 +16,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include <jpeglib.h>
+#include <jconfigint.h>
 
 #ifndef IMAGE_PACK_INIT_EXPORT
 #if defined(_WIN32)
@@ -39,6 +40,14 @@
 #define FALSE 0
 #endif
 
+#if defined(_MSC_VER) && !defined(__clang__)
+#define IP_RESTRICT __restrict
+#elif defined(__GNUC__) || defined(__clang__)
+#define IP_RESTRICT __restrict
+#else
+#define IP_RESTRICT
+#endif
+
 typedef enum { IP_ALGO_JPEG_TURBO = 1, IP_ALGO_MOZJPEG = 2 } ip_algo_t;
 
 typedef enum {
@@ -105,6 +114,7 @@ typedef struct {
     int max_width;
     int max_height;
     size_t max_output_size;
+    size_t max_input_size;
 
     ip_status_t status;
     char error_message[512];
@@ -117,6 +127,19 @@ typedef struct {
 
     unsigned char *scratch_row;
     size_t scratch_row_size;
+
+    struct {
+        int marker;
+        unsigned char *data;
+        unsigned int len;
+    } *preserved_markers;
+    size_t preserved_marker_count;
+    size_t preserved_marker_capacity;
+
+    unsigned char *transient_jpeg_buf;
+    unsigned char *transient_decode_buf;
+    unsigned char *transient_scratch_buf;
+    int source_orientation;
 } ip_context_t;
 
 typedef struct {
@@ -152,6 +175,7 @@ static ID id_max_pixels;
 static ID id_max_width;
 static ID id_max_height;
 static ID id_max_output_size;
+static ID id_max_input_size;
 
 static ip_context_t *ip_context_new(void);
 static void ip_context_free(ip_context_t *ctx);
@@ -187,13 +211,45 @@ static void validate_limits_for_pixels(ip_context_t *ctx);
 
 static int ip_jpeg_decode_to_pixels(ip_context_t *ctx, unsigned char **pixels, int *width,
                                     int *height, int *channels, int fast_decode_mode);
+static int ip_decode_jpeg_to_luma_buffer(ip_context_t *ctx, const unsigned char *data, size_t size,
+                                         unsigned char **luma, int *width, int *height);
 static int guarded_compress_jpeg_input_with_mode(ip_context_t *ctx, int mozjpeg_size_mode);
 static int ip_jpeg_turbo_compress(ip_context_t *ctx);
 static int ip_mozjpeg_compress(ip_context_t *ctx);
+static int ip_lossless_optimize_jpeg(ip_context_t *ctx);
+static int ip_run_optimize_context(ip_context_t *ctx);
 
-#if defined(IMAGE_PACK_ENABLE_LOSSLESS_TRANSCODE_FAST_PATH)
-static int ip_jpeg_transcode_coefficients(ip_context_t *ctx, int mozjpeg_size_mode);
-#endif
+typedef struct {
+    VALUE self, input, input_kind, output, output_kind, algo, quality, min_ssim;
+    VALUE mozjpeg_trellis, progressive, strip_metadata, execution, cancellable, has_scheduler;
+    ip_context_t *ctx;
+} ip_compress_jpeg_call_t;
+
+typedef struct {
+    VALUE self, buffer, width, height, channels, output, output_kind, algo, quality, min_ssim;
+    VALUE progressive, execution, cancellable, has_scheduler;
+    ip_context_t *ctx;
+} ip_compress_pixels_call_t;
+
+typedef struct {
+    VALUE self, input, input_kind, output, output_kind, progressive, strip_metadata;
+    VALUE execution, cancellable, has_scheduler;
+    ip_context_t *ctx;
+} ip_optimize_jpeg_call_t;
+
+typedef struct {
+    VALUE self, input, input_kind;
+    ip_context_t *ctx;
+} ip_inspect_call_t;
+
+static VALUE ip_call_cleanup(VALUE ptr) {
+    ip_context_t **ctx_ptr = (ip_context_t **)ptr;
+    if (ctx_ptr && *ctx_ptr) {
+        ip_context_free(*ctx_ptr);
+        *ctx_ptr = NULL;
+    }
+    return Qnil;
+}
 
 static VALUE ip_compress_jpeg_entry(VALUE self, VALUE input, VALUE input_kind, VALUE output,
                                     VALUE output_kind, VALUE algo, VALUE quality, VALUE min_ssim,
@@ -201,8 +257,11 @@ static VALUE ip_compress_jpeg_entry(VALUE self, VALUE input, VALUE input_kind, V
                                     VALUE execution, VALUE cancellable, VALUE has_scheduler);
 static VALUE ip_compress_pixels_entry(VALUE self, VALUE buffer, VALUE width, VALUE height,
                                       VALUE channels, VALUE output, VALUE output_kind, VALUE algo,
-                                      VALUE quality, VALUE progressive, VALUE execution,
-                                      VALUE cancellable, VALUE has_scheduler);
+                                      VALUE quality, VALUE min_ssim, VALUE progressive,
+                                      VALUE execution, VALUE cancellable, VALUE has_scheduler);
+static VALUE ip_optimize_jpeg_entry(VALUE self, VALUE input, VALUE input_kind, VALUE output,
+                                    VALUE output_kind, VALUE progressive, VALUE strip_metadata,
+                                    VALUE execution, VALUE cancellable, VALUE has_scheduler);
 
 static VALUE ip_status_to_exception(ip_status_t status) {
     switch (status) {
@@ -251,12 +310,15 @@ static int ip_checked_image_size(int width, int height, int channels, size_t *ou
     return ip_checked_mul_size(pixels, (size_t)channels, out);
 }
 
+static void *ip_malloc_hot(size_t size) {
+    return malloc(size);
+}
+
 static void ip_validate_quality_or_raise(ip_context_t *ctx) {
     if (ctx->quality >= 1 && ctx->quality <= 100)
         return;
 
     int quality = ctx->quality;
-    ip_context_free(ctx);
     rb_raise(rb_eImagePackInvalidArgumentError, "quality must be Integer 1..100, got: %d", quality);
 }
 
@@ -267,7 +329,6 @@ static void ip_validate_min_ssim_or_raise(ip_context_t *ctx) {
         return;
 
     double min_ssim = ctx->min_ssim;
-    ip_context_free(ctx);
     rb_raise(rb_eImagePackInvalidArgumentError,
              "min_ssim must be Numeric > 0.0 and <= 1.0, got: %.17g", min_ssim);
 }
@@ -297,6 +358,8 @@ static ip_context_t *ip_context_new(void) {
     ctx->max_width = 30000;
     ctx->max_height = 30000;
     ctx->max_output_size = 256 * 1024 * 1024;
+    ctx->max_input_size = 256 * 1024 * 1024;
+    ctx->source_orientation = 1;
     atomic_init(&ctx->cancelled, 0);
     return ctx;
 }
@@ -309,6 +372,16 @@ static void ip_context_free(ip_context_t *ctx) {
     free(ctx->owned_pixel_data);
     free(ctx->output_path);
     free(ctx->scratch_row);
+    free(ctx->transient_jpeg_buf);
+    free(ctx->transient_decode_buf);
+    free(ctx->transient_scratch_buf);
+
+    if (ctx->preserved_markers) {
+        for (size_t i = 0; i < ctx->preserved_marker_count; i++) {
+            free(ctx->preserved_markers[i].data);
+        }
+        free(ctx->preserved_markers);
+    }
 
     if (ctx->output_data && ctx->output_owner == IP_OUTPUT_OWNER_MALLOC) {
         free(ctx->output_data);
@@ -405,6 +478,12 @@ static int read_file_to_owned_buffer(ip_context_t *ctx, const char *path) {
     }
     rewind(fp);
 
+    if (ctx->max_input_size > 0 && (size_t)size > ctx->max_input_size) {
+        fclose(fp);
+        ip_context_set_error(ctx, IP_ERR_LIMIT, "input file exceeds max_input_size");
+        return 0;
+    }
+
     unsigned char *data = (unsigned char *)malloc((size_t)size);
     if (!data && size > 0) {
         fclose(fp);
@@ -435,6 +514,16 @@ static int ip_prepare_input_bytes(ip_context_t *ctx, VALUE input, ip_input_kind_
     if (kind == IP_INPUT_BYTES) {
         Check_Type(input, T_STRING);
         size_t len = (size_t)RSTRING_LEN(input);
+        if (ctx->max_input_size > 0 && len > ctx->max_input_size) {
+            ip_context_set_error(ctx, IP_ERR_LIMIT, "input bytes exceed max_input_size");
+            return 0;
+        }
+        if (ctx->requested_execution == IP_EXEC_DIRECT) {
+            ctx->input_data = (const unsigned char *)RSTRING_PTR(input);
+            ctx->input_size = len;
+            return 1;
+        }
+
         unsigned char *copy = (unsigned char *)malloc(len);
         if (!copy && len > 0) {
             ip_context_set_error(ctx, IP_ERR_OOM, "failed to copy binary String input");
@@ -452,6 +541,10 @@ static int ip_prepare_input_bytes(ip_context_t *ctx, VALUE input, ip_input_kind_
         VALUE str = io_buffer_to_string(input);
         StringValue(str);
         size_t len = (size_t)RSTRING_LEN(str);
+        if (ctx->max_input_size > 0 && len > ctx->max_input_size) {
+            ip_context_set_error(ctx, IP_ERR_LIMIT, "input IO::Buffer exceeds max_input_size");
+            return 0;
+        }
         unsigned char *copy = (unsigned char *)malloc(len);
         if (!copy && len > 0) {
             ip_context_set_error(ctx, IP_ERR_OOM, "failed to copy IO::Buffer input");
@@ -492,7 +585,18 @@ static int ip_prepare_pixels(ip_context_t *ctx, VALUE buffer, int width, int hei
         return 0;
     }
 
-    unsigned char *copy = (unsigned char *)malloc(expected);
+    if (ctx->requested_execution == IP_EXEC_DIRECT && RB_TYPE_P(buffer, T_STRING)) {
+        ctx->pixel_data = (const unsigned char *)RSTRING_PTR(str);
+        ctx->pixel_size = expected;
+        ctx->width = width;
+        ctx->height = height;
+        ctx->channels = channels;
+        ctx->bit_depth = 8;
+        ctx->decoded_bytes = expected;
+        return 1;
+    }
+
+    unsigned char *copy = (unsigned char *)ip_malloc_hot(expected);
     if (!copy && expected > 0) {
         ip_context_set_error(ctx, IP_ERR_OOM, "failed to copy pixel buffer");
         return 0;
@@ -530,26 +634,248 @@ static VALUE ip_finish_output(ip_context_t *ctx, ip_output_kind_t kind) {
     ip_raise_for_status(ctx);
 
     if (kind == IP_OUTPUT_RETURN_STRING) {
+        if (ctx->output_size > (size_t)LONG_MAX)
+            rb_raise(rb_eImagePackLimitExceededError, "output is too large for a Ruby String");
         VALUE out = rb_str_new((const char *)ctx->output_data, (long)ctx->output_size);
         rb_enc_associate(out, rb_ascii8bit_encoding());
         return out;
     }
 
-    FILE *fp = fopen(ctx->output_path, "wb");
+    size_t path_len = strlen(ctx->output_path);
+    char suffix[64];
+    snprintf(suffix, sizeof(suffix), ".tmp.image_pack.%p", (void *)ctx);
+    size_t tmp_len = path_len + strlen(suffix) + 1;
+    char *tmp_path = (char *)malloc(tmp_len);
+    if (!tmp_path)
+        rb_raise(rb_eImagePackOutOfMemoryError, "failed to allocate temporary output path");
+    snprintf(tmp_path, tmp_len, "%s%s", ctx->output_path, suffix);
+
+    FILE *fp = fopen(tmp_path, "wb");
     if (!fp) {
+        free(tmp_path);
         rb_raise(rb_eImagePackInvalidArgumentError, "failed to open output path: %s",
                  ctx->output_path);
     }
 
     size_t written = fwrite(ctx->output_data, 1, ctx->output_size, fp);
-    fclose(fp);
-    if (written != ctx->output_size) {
+    int write_failed = written != ctx->output_size || ferror(fp);
+    int close_failed = fclose(fp) != 0;
+    if (write_failed || close_failed) {
+        remove(tmp_path);
+        free(tmp_path);
         rb_raise(rb_eImagePackEncodeError, "failed to write full JPEG output");
     }
 
+    if (rename(tmp_path, ctx->output_path) != 0) {
+        remove(tmp_path);
+        free(tmp_path);
+        rb_raise(rb_eImagePackEncodeError, "failed to move temporary JPEG output into place");
+    }
+
+    free(tmp_path);
     return Qtrue;
 }
 
+static int ip_save_marker(ip_context_t *ctx, int marker, const unsigned char *data,
+                          unsigned int len) {
+    if (ctx->preserved_marker_count == ctx->preserved_marker_capacity) {
+        size_t new_cap =
+            ctx->preserved_marker_capacity == 0 ? 4 : ctx->preserved_marker_capacity * 2;
+        void *new_buf = realloc(ctx->preserved_markers, new_cap * sizeof(*ctx->preserved_markers));
+        if (!new_buf)
+            return 0;
+        ctx->preserved_markers = new_buf;
+        ctx->preserved_marker_capacity = new_cap;
+    }
+
+    unsigned char *copy = (unsigned char *)malloc(len);
+    if (!copy && len > 0)
+        return 0;
+    if (len > 0)
+        memcpy(copy, data, len);
+
+    ctx->preserved_markers[ctx->preserved_marker_count].marker = marker;
+    ctx->preserved_markers[ctx->preserved_marker_count].data = copy;
+    ctx->preserved_markers[ctx->preserved_marker_count].len = len;
+    ctx->preserved_marker_count++;
+    return 1;
+}
+
+static int ip_save_markers_from_decompress(ip_context_t *ctx,
+                                           struct jpeg_decompress_struct *cinfo) {
+    jpeg_saved_marker_ptr m;
+    for (m = cinfo->marker_list; m != NULL; m = m->next) {
+        if (m->marker == (JPEG_APP0 + 0))
+            continue;
+
+        if (!ip_save_marker(ctx, m->marker, m->data, m->data_length)) {
+            ip_context_set_error(ctx, IP_ERR_OOM, "failed to preserve JPEG metadata marker");
+            return 0;
+        }
+    }
+    return 1;
+}
+
+static uint16_t ip_exif_u16(const unsigned char *p, int little_endian) {
+    if (little_endian)
+        return (uint16_t)p[0] | ((uint16_t)p[1] << 8);
+    return ((uint16_t)p[0] << 8) | (uint16_t)p[1];
+}
+
+static uint32_t ip_exif_u32(const unsigned char *p, int little_endian) {
+    if (little_endian)
+        return (uint32_t)p[0] | ((uint32_t)p[1] << 8) | ((uint32_t)p[2] << 16) |
+               ((uint32_t)p[3] << 24);
+    return ((uint32_t)p[0] << 24) | ((uint32_t)p[1] << 16) | ((uint32_t)p[2] << 8) | (uint32_t)p[3];
+}
+
+static int ip_parse_exif_orientation(const unsigned char *data, unsigned int len) {
+    if (!data || len < 14)
+        return 1;
+    if (memcmp(data, "Exif\0\0", 6) != 0)
+        return 1;
+
+    const unsigned char *tiff = data + 6;
+    size_t tiff_len = (size_t)len - 6;
+    if (tiff_len < 8)
+        return 1;
+
+    int little_endian = 0;
+    if (tiff[0] == 'I' && tiff[1] == 'I')
+        little_endian = 1;
+    else if (tiff[0] == 'M' && tiff[1] == 'M')
+        little_endian = 0;
+    else
+        return 1;
+
+    if (ip_exif_u16(tiff + 2, little_endian) != 42)
+        return 1;
+
+    uint32_t ifd0_offset = ip_exif_u32(tiff + 4, little_endian);
+    if (ifd0_offset > tiff_len || tiff_len - ifd0_offset < 2)
+        return 1;
+
+    const unsigned char *ifd = tiff + ifd0_offset;
+    size_t ifd_len = tiff_len - ifd0_offset;
+    uint16_t entries = ip_exif_u16(ifd, little_endian);
+    ifd += 2;
+    ifd_len -= 2;
+
+    for (uint16_t i = 0; i < entries; i++) {
+        if (ifd_len < 12)
+            return 1;
+
+        uint16_t tag = ip_exif_u16(ifd, little_endian);
+        uint16_t type = ip_exif_u16(ifd + 2, little_endian);
+        uint32_t count = ip_exif_u32(ifd + 4, little_endian);
+
+        if (tag == 0x0112 && type == 3 && count == 1) {
+            uint16_t orientation = ip_exif_u16(ifd + 8, little_endian);
+            if (orientation >= 1 && orientation <= 8)
+                return (int)orientation;
+            return 1;
+        }
+
+        ifd += 12;
+        ifd_len -= 12;
+    }
+
+    return 1;
+}
+
+static int ip_read_exif_orientation_from_decompress(struct jpeg_decompress_struct *cinfo) {
+    jpeg_saved_marker_ptr m;
+    for (m = cinfo->marker_list; m != NULL; m = m->next) {
+        if (m->marker == (JPEG_APP0 + 1)) {
+            int orientation = ip_parse_exif_orientation(m->data, m->data_length);
+            if (orientation >= 2 && orientation <= 8)
+                return orientation;
+        }
+    }
+    return 1;
+}
+
+static int ip_transform_pixels_for_orientation(ip_context_t *ctx, unsigned char **pixels,
+                                               int *width, int *height, int channels) {
+    int orientation = ctx->source_orientation;
+    if (orientation <= 1 || orientation > 8)
+        return 1;
+
+    int src_w = *width;
+    int src_h = *height;
+    int dst_w = (orientation >= 5 && orientation <= 8) ? src_h : src_w;
+    int dst_h = (orientation >= 5 && orientation <= 8) ? src_w : src_h;
+    size_t out_size = 0;
+    if (!ip_checked_image_size(dst_w, dst_h, channels, &out_size)) {
+        ip_context_set_error(ctx, IP_ERR_LIMIT, "oriented image size overflows native size");
+        return 0;
+    }
+
+    unsigned char *src = *pixels;
+    unsigned char *dst = (unsigned char *)ip_malloc_hot(out_size);
+    if (!dst && out_size > 0) {
+        ip_context_set_error(ctx, IP_ERR_OOM, "failed to allocate EXIF-oriented pixel buffer");
+        return 0;
+    }
+
+    for (int y = 0; y < src_h; y++) {
+        for (int x = 0; x < src_w; x++) {
+            int dx = x;
+            int dy = y;
+            switch (orientation) {
+            case 2:
+                dx = src_w - 1 - x;
+                dy = y;
+                break;
+            case 3:
+                dx = src_w - 1 - x;
+                dy = src_h - 1 - y;
+                break;
+            case 4:
+                dx = x;
+                dy = src_h - 1 - y;
+                break;
+            case 5:
+                dx = y;
+                dy = x;
+                break;
+            case 6:
+                dx = src_h - 1 - y;
+                dy = x;
+                break;
+            case 7:
+                dx = src_h - 1 - y;
+                dy = src_w - 1 - x;
+                break;
+            case 8:
+                dx = y;
+                dy = src_w - 1 - x;
+                break;
+            default:
+                break;
+            }
+
+            memcpy(dst + (((size_t)dy * (size_t)dst_w + (size_t)dx) * (size_t)channels),
+                   src + (((size_t)y * (size_t)src_w + (size_t)x) * (size_t)channels),
+                   (size_t)channels);
+        }
+    }
+
+    free(src);
+    *pixels = dst;
+    *width = dst_w;
+    *height = dst_h;
+    return 1;
+}
+
+static void ip_write_preserved_markers(ip_context_t *ctx, struct jpeg_compress_struct *cinfo) {
+    for (size_t i = 0; i < ctx->preserved_marker_count; i++) {
+        jpeg_write_marker(cinfo, ctx->preserved_markers[i].marker,
+                          (const JOCTET *)ctx->preserved_markers[i].data,
+                          ctx->preserved_markers[i].len);
+    }
+}
+
 static void ip_jpeg_invalid_error_exit(j_common_ptr cinfo) {
     ip_jpeg_error_mgr *err = (ip_jpeg_error_mgr *)cinfo->err;
     char buffer[JMSG_LENGTH_MAX];
@@ -601,11 +927,23 @@ static int ip_inspect_jpeg_header(ip_context_t *ctx) {
         return 0;
     }
 
+    if (cinfo.num_components == 4 || cinfo.jpeg_color_space == JCS_CMYK ||
+        cinfo.jpeg_color_space == JCS_YCCK) {
+        jpeg_destroy_decompress(&cinfo);
+        ctx->jmp_armed = 0;
+        ip_context_set_error(ctx, IP_ERR_UNSUPPORTED,
+                             "CMYK/YCCK JPEG input is not supported in this release");
+        return 0;
+    }
+
     ctx->width = (int)cinfo.image_width;
     ctx->height = (int)cinfo.image_height;
     ctx->channels = cinfo.num_components;
     if (ctx->channels != 1 && ctx->channels != 3 && ctx->channels != 4) {
-        ctx->channels = 3;
+        jpeg_destroy_decompress(&cinfo);
+        ctx->jmp_armed = 0;
+        ip_context_set_error(ctx, IP_ERR_UNSUPPORTED, "JPEG component count is not supported");
+        return 0;
     }
     ctx->bit_depth = 8;
     ctx->decoded_bytes = (size_t)ctx->width * (size_t)ctx->height * (size_t)ctx->channels;
@@ -618,15 +956,15 @@ static int ip_inspect_jpeg_header(ip_context_t *ctx) {
         return 0;
     }
 
-    if (ctx->width > ctx->max_width) {
+    if (ctx->max_width > 0 && ctx->width > ctx->max_width) {
         ip_context_set_error(ctx, IP_ERR_LIMIT, "image width exceeds max_width");
         return 0;
     }
-    if (ctx->height > ctx->max_height) {
+    if (ctx->max_height > 0 && ctx->height > ctx->max_height) {
         ip_context_set_error(ctx, IP_ERR_LIMIT, "image height exceeds max_height");
         return 0;
     }
-    if ((uint64_t)ctx->width * (uint64_t)ctx->height > ctx->max_pixels) {
+    if (ctx->max_pixels > 0 && (uint64_t)ctx->width * (uint64_t)ctx->height > ctx->max_pixels) {
         ip_context_set_error(ctx, IP_ERR_LIMIT, "image pixels exceed max_pixels");
         return 0;
     }
@@ -634,32 +972,43 @@ static int ip_inspect_jpeg_header(ip_context_t *ctx) {
     return 1;
 }
 
-static VALUE ip_inspect_image_entry(VALUE self, VALUE input, VALUE input_kind) {
-    (void)self;
+static VALUE ip_inspect_image_entry_body(VALUE ptr) {
+    ip_inspect_call_t *call = (ip_inspect_call_t *)ptr;
     ip_context_t *ctx = ip_context_new();
     if (!ctx)
         rb_raise(rb_eImagePackOutOfMemoryError, "failed to allocate native context");
+    call->ctx = ctx;
 
-    if (!ip_prepare_input_bytes(ctx, input, ip_parse_input_kind(input_kind)) ||
+    if (!ip_prepare_input_bytes(ctx, call->input, ip_parse_input_kind(call->input_kind)) ||
         !ip_inspect_jpeg_header(ctx)) {
-        VALUE exception = ip_status_to_exception(ctx->status);
-        char message[512];
-        snprintf(message, sizeof(message), "%s", ctx->error_message);
-        ip_context_free(ctx);
-        rb_raise(exception, "%s", message[0] ? message : "failed to inspect JPEG image");
+        ip_raise_for_status(ctx);
+        rb_raise(rb_eImagePackInvalidImageError, "failed to inspect JPEG image");
     }
 
+    int width = ctx->width;
+    int height = ctx->height;
+    int channels = ctx->channels;
+    int bit_depth = ctx->bit_depth;
+    size_t decoded_bytes = ctx->decoded_bytes;
+
+    ip_context_free(ctx);
+    call->ctx = NULL;
+
     VALUE hash = rb_hash_new();
     rb_hash_aset(hash, ID2SYM(rb_intern("format")), ID2SYM(rb_intern("jpeg")));
-    rb_hash_aset(hash, ID2SYM(rb_intern("width")), INT2NUM(ctx->width));
-    rb_hash_aset(hash, ID2SYM(rb_intern("height")), INT2NUM(ctx->height));
-    rb_hash_aset(hash, ID2SYM(rb_intern("channels")), INT2NUM(ctx->channels));
-    rb_hash_aset(hash, ID2SYM(rb_intern("bit_depth")), INT2NUM(ctx->bit_depth));
-    rb_hash_aset(hash, ID2SYM(rb_intern("decoded_bytes")), SIZET2NUM(ctx->decoded_bytes));
-    ip_context_free(ctx);
+    rb_hash_aset(hash, ID2SYM(rb_intern("width")), INT2NUM(width));
+    rb_hash_aset(hash, ID2SYM(rb_intern("height")), INT2NUM(height));
+    rb_hash_aset(hash, ID2SYM(rb_intern("channels")), INT2NUM(channels));
+    rb_hash_aset(hash, ID2SYM(rb_intern("bit_depth")), INT2NUM(bit_depth));
+    rb_hash_aset(hash, ID2SYM(rb_intern("decoded_bytes")), SIZET2NUM(decoded_bytes));
     return hash;
 }
 
+static VALUE ip_inspect_image_entry(VALUE self, VALUE input, VALUE input_kind) {
+    ip_inspect_call_t call = {self, input, input_kind, NULL};
+    return rb_ensure(ip_inspect_image_entry_body, (VALUE)&call, ip_call_cleanup, (VALUE)&call.ctx);
+}
+
 static J_COLOR_SPACE color_space_for_channels(int channels) {
     return channels == 1 ? JCS_GRAYSCALE : JCS_RGB;
 }
@@ -677,6 +1026,15 @@ static void configure_mozjpeg_features_after_defaults(struct jpeg_compress_struc
     if (mozjpeg_size_mode) {
         cinfo->optimize_coding = TRUE;
 
+        if (progressive_requested) {
+            jpeg_simple_progression(cinfo);
+            jpeg_c_set_bool_param(cinfo, JBOOLEAN_OPTIMIZE_SCANS, TRUE);
+        } else {
+            cinfo->scan_info = NULL;
+            cinfo->num_scans = 0;
+            jpeg_c_set_bool_param(cinfo, JBOOLEAN_OPTIMIZE_SCANS, FALSE);
+        }
+
         if (!mozjpeg_trellis_enabled) {
             jpeg_c_set_bool_param(cinfo, JBOOLEAN_TRELLIS_QUANT, FALSE);
             jpeg_c_set_bool_param(cinfo, JBOOLEAN_TRELLIS_QUANT_DC, FALSE);
@@ -710,63 +1068,77 @@ static void configure_mozjpeg_features_after_defaults(struct jpeg_compress_struc
     }
 }
 
-static int prepare_encode_row(ip_context_t *ctx, JDIMENSION y, JSAMPROW *row) {
-    if (ctx->channels == 4) {
-        size_t rgb_row_size = 0;
-        if (!ip_checked_mul_size((size_t)ctx->width, 3, &rgb_row_size)) {
-            ip_context_set_error(ctx, IP_ERR_LIMIT, "RGBA scratch row size overflow");
-            return 0;
+static int prepare_encode_rows(ip_context_t *ctx, JDIMENSION start, JDIMENSION batch,
+                               JSAMPROW *rows) {
+    if (ctx->channels != 4) {
+        for (JDIMENSION i = 0; i < batch; i++) {
+            JDIMENSION y = start + i;
+            rows[i] = (JSAMPROW)(ctx->pixel_data +
+                                 ((size_t)y * (size_t)ctx->width * (size_t)ctx->channels));
         }
+        return 1;
+    }
 
-        if (ctx->scratch_row_size < rgb_row_size) {
-            unsigned char *new_row = (unsigned char *)realloc(ctx->scratch_row, rgb_row_size);
-            if (!new_row) {
-                ip_context_set_error(ctx, IP_ERR_OOM, "failed to allocate RGBA scratch row");
-                return 0;
-            }
-            ctx->scratch_row = new_row;
-            ctx->scratch_row_size = rgb_row_size;
+    size_t rgb_row_size = 0;
+    size_t scratch_size = 0;
+    if (!ip_checked_mul_size((size_t)ctx->width, 3, &rgb_row_size) ||
+        !ip_checked_mul_size(rgb_row_size, (size_t)batch, &scratch_size)) {
+        ip_context_set_error(ctx, IP_ERR_LIMIT, "RGBA scratch row size overflow");
+        return 0;
+    }
+
+    if (ctx->scratch_row_size < scratch_size) {
+        unsigned char *new_rows = (unsigned char *)realloc(ctx->scratch_row, scratch_size);
+        if (!new_rows) {
+            ip_context_set_error(ctx, IP_ERR_OOM, "failed to allocate RGBA scratch rows");
+            return 0;
         }
+        ctx->scratch_row = new_rows;
+        ctx->scratch_row_size = scratch_size;
+    }
 
-        const unsigned char *src = ctx->pixel_data + ((size_t)y * (size_t)ctx->width * 4);
-        unsigned char *dst = ctx->scratch_row;
-        for (int x = 0; x < ctx->width; x++) {
+    for (JDIMENSION i = 0; i < batch; i++) {
+        JDIMENSION y = start + i;
+        const unsigned char *IP_RESTRICT src =
+            ctx->pixel_data + ((size_t)y * (size_t)ctx->width * 4);
+        unsigned char *IP_RESTRICT dst = ctx->scratch_row + ((size_t)i * rgb_row_size);
+        const int w = ctx->width;
+        for (int x = 0; x < w; x++) {
             dst[x * 3 + 0] = src[x * 4 + 0];
             dst[x * 3 + 1] = src[x * 4 + 1];
             dst[x * 3 + 2] = src[x * 4 + 2];
         }
-        *row = ctx->scratch_row;
-        return 1;
+        rows[i] = dst;
     }
 
-    *row = (JSAMPROW)(ctx->pixel_data + ((size_t)y * (size_t)ctx->width * (size_t)ctx->channels));
     return 1;
 }
 
 static int encode_pixels_with_libjpeg(ip_context_t *ctx, int mozjpeg_size_mode) {
     struct jpeg_compress_struct cinfo;
     ip_jpeg_error_mgr jerr;
+    unsigned long jpeg_size = 0;
     memset(&cinfo, 0, sizeof(cinfo));
     memset(&jerr, 0, sizeof(jerr));
 
     cinfo.err = jpeg_std_error(&jerr.pub);
     jerr.pub.error_exit = ip_jpeg_encode_error_exit;
     jerr.ctx = ctx;
+    ctx->transient_jpeg_buf = NULL;
 
     ctx->jmp_armed = 1;
     if (setjmp(ctx->jmpbuf)) {
         ctx->jmp_armed = 0;
         jpeg_destroy_compress(&cinfo);
+        free(ctx->transient_jpeg_buf);
+        ctx->transient_jpeg_buf = NULL;
         if (ctx->status == IP_OK)
             ip_context_set_error(ctx, IP_ERR_ENCODE, "JPEG encode failed");
         return 0;
     }
 
     jpeg_create_compress(&cinfo);
-
-    unsigned char *jpeg_buf = NULL;
-    unsigned long jpeg_size = 0;
-    jpeg_mem_dest(&cinfo, &jpeg_buf, &jpeg_size);
+    jpeg_mem_dest(&cinfo, &ctx->transient_jpeg_buf, &jpeg_size);
 
     cinfo.image_width = (JDIMENSION)ctx->width;
     cinfo.image_height = (JDIMENSION)ctx->height;
@@ -781,40 +1153,52 @@ static int encode_pixels_with_libjpeg(ip_context_t *ctx, int mozjpeg_size_mode)
 
     jpeg_start_compress(&cinfo, TRUE);
 
+    if (!ctx->strip_metadata) {
+        ip_write_preserved_markers(ctx, &cinfo);
+    }
+
     while (cinfo.next_scanline < cinfo.image_height) {
-        if (ctx->cancellable_requested && (cinfo.next_scanline % 16 == 0) &&
-            atomic_load(&ctx->cancelled)) {
+        if (ctx->cancellable_requested && atomic_load(&ctx->cancelled)) {
             ip_context_set_error(ctx, IP_ERR_CANCELLED, "JPEG encode cancelled");
             jpeg_abort_compress(&cinfo);
             jpeg_destroy_compress(&cinfo);
-            free(jpeg_buf);
+            free(ctx->transient_jpeg_buf);
+            ctx->transient_jpeg_buf = NULL;
             ctx->jmp_armed = 0;
             return 0;
         }
 
-        JSAMPROW row = NULL;
-        if (!prepare_encode_row(ctx, cinfo.next_scanline, &row)) {
+        JSAMPROW rows[16];
+        JDIMENSION start_scanline = cinfo.next_scanline;
+        JDIMENSION batch = cinfo.image_height - start_scanline;
+        if (batch > 16)
+            batch = 16;
+
+        if (!prepare_encode_rows(ctx, start_scanline, batch, rows)) {
             jpeg_abort_compress(&cinfo);
             jpeg_destroy_compress(&cinfo);
-            free(jpeg_buf);
+            free(ctx->transient_jpeg_buf);
+            ctx->transient_jpeg_buf = NULL;
             ctx->jmp_armed = 0;
             return 0;
         }
 
-        jpeg_write_scanlines(&cinfo, &row, 1);
+        jpeg_write_scanlines(&cinfo, rows, batch);
     }
 
     jpeg_finish_compress(&cinfo);
     jpeg_destroy_compress(&cinfo);
     ctx->jmp_armed = 0;
 
-    if ((size_t)jpeg_size > ctx->max_output_size) {
-        free(jpeg_buf);
+    if (ctx->max_output_size > 0 && (size_t)jpeg_size > ctx->max_output_size) {
+        free(ctx->transient_jpeg_buf);
+        ctx->transient_jpeg_buf = NULL;
         ip_context_set_error(ctx, IP_ERR_LIMIT, "output exceeds max_output_size");
         return 0;
     }
 
-    ctx->output_data = jpeg_buf;
+    ctx->output_data = ctx->transient_jpeg_buf;
+    ctx->transient_jpeg_buf = NULL;
     ctx->output_size = (size_t)jpeg_size;
     ctx->output_capacity = (size_t)jpeg_size;
     ctx->output_owner = IP_OUTPUT_OWNER_MALLOC;
@@ -831,11 +1215,15 @@ static int ip_jpeg_decode_to_pixels(ip_context_t *ctx, unsigned char **pixels, i
     cinfo.err = jpeg_std_error(&jerr.pub);
     jerr.pub.error_exit = ip_jpeg_invalid_error_exit;
     jerr.ctx = ctx;
+    ctx->transient_decode_buf = NULL;
+    ctx->source_orientation = 1;
 
     ctx->jmp_armed = 1;
     if (setjmp(ctx->jmpbuf)) {
         ctx->jmp_armed = 0;
         jpeg_destroy_decompress(&cinfo);
+        free(ctx->transient_decode_buf);
+        ctx->transient_decode_buf = NULL;
         if (ctx->status == IP_OK)
             ip_context_set_error(ctx, IP_ERR_INVALID_IMAGE, "JPEG decode failed");
         return 0;
@@ -843,6 +1231,15 @@ static int ip_jpeg_decode_to_pixels(ip_context_t *ctx, unsigned char **pixels, i
 
     jpeg_create_decompress(&cinfo);
     jpeg_mem_src(&cinfo, ctx->input_data, (unsigned long)ctx->input_size);
+
+    jpeg_save_markers(&cinfo, JPEG_APP0 + 1, 0xFFFF);
+    if (!ctx->strip_metadata) {
+        jpeg_save_markers(&cinfo, JPEG_COM, 0xFFFF);
+        for (int app = 2; app < 16; app++) {
+            jpeg_save_markers(&cinfo, JPEG_APP0 + app, 0xFFFF);
+        }
+    }
+
     int rc = jpeg_read_header(&cinfo, TRUE);
     if (rc != JPEG_HEADER_OK) {
         jpeg_destroy_decompress(&cinfo);
@@ -851,6 +1248,8 @@ static int ip_jpeg_decode_to_pixels(ip_context_t *ctx, unsigned char **pixels, i
         return 0;
     }
 
+    ctx->source_orientation = ip_read_exif_orientation_from_decompress(&cinfo);
+
     if (cinfo.image_width > (JDIMENSION)INT_MAX || cinfo.image_height > (JDIMENSION)INT_MAX) {
         jpeg_destroy_decompress(&cinfo);
         ctx->jmp_armed = 0;
@@ -858,6 +1257,15 @@ static int ip_jpeg_decode_to_pixels(ip_context_t *ctx, unsigned char **pixels, i
         return 0;
     }
 
+    if (cinfo.num_components == 4 || cinfo.jpeg_color_space == JCS_CMYK ||
+        cinfo.jpeg_color_space == JCS_YCCK) {
+        jpeg_destroy_decompress(&cinfo);
+        ctx->jmp_armed = 0;
+        ip_context_set_error(ctx, IP_ERR_UNSUPPORTED,
+                             "CMYK/YCCK JPEG input is not supported in this release");
+        return 0;
+    }
+
     int ch = cinfo.num_components == 1 ? 1 : 3;
 
     ctx->width = (int)cinfo.image_width;
@@ -892,6 +1300,15 @@ static int ip_jpeg_decode_to_pixels(ip_context_t *ctx, unsigned char **pixels, i
     }
 
     jpeg_start_decompress(&cinfo);
+
+    if (!ctx->strip_metadata) {
+        if (!ip_save_markers_from_decompress(ctx, &cinfo)) {
+            jpeg_destroy_decompress(&cinfo);
+            ctx->jmp_armed = 0;
+            return 0;
+        }
+    }
+
     size_t row_stride = 0;
     size_t size = 0;
     if (!ip_checked_mul_size((size_t)cinfo.output_width, (size_t)cinfo.output_components,
@@ -902,8 +1319,8 @@ static int ip_jpeg_decode_to_pixels(ip_context_t *ctx, unsigned char **pixels, i
         ip_context_set_error(ctx, IP_ERR_LIMIT, "decoded image buffer size overflow");
         return 0;
     }
-    unsigned char *buf = (unsigned char *)malloc(size);
-    if (!buf && size > 0) {
+    ctx->transient_decode_buf = (unsigned char *)ip_malloc_hot(size);
+    if (!ctx->transient_decode_buf && size > 0) {
         jpeg_destroy_decompress(&cinfo);
         ctx->jmp_armed = 0;
         ip_context_set_error(ctx, IP_ERR_OOM, "failed to allocate decoded pixel buffer");
@@ -911,8 +1328,27 @@ static int ip_jpeg_decode_to_pixels(ip_context_t *ctx, unsigned char **pixels, i
     }
 
     while (cinfo.output_scanline < cinfo.output_height) {
-        JSAMPROW row = buf + ((size_t)cinfo.output_scanline * row_stride);
-        jpeg_read_scanlines(&cinfo, &row, 1);
+        if (ctx->cancellable_requested && atomic_load(&ctx->cancelled)) {
+            ip_context_set_error(ctx, IP_ERR_CANCELLED, "JPEG decode cancelled");
+            jpeg_abort_decompress(&cinfo);
+            jpeg_destroy_decompress(&cinfo);
+            free(ctx->transient_decode_buf);
+            ctx->transient_decode_buf = NULL;
+            ctx->jmp_armed = 0;
+            return 0;
+        }
+
+        JSAMPROW rows[16];
+        JDIMENSION batch = cinfo.output_height - cinfo.output_scanline;
+        if (batch > 16)
+            batch = 16;
+
+        for (JDIMENSION i = 0; i < batch; i++) {
+            rows[i] =
+                ctx->transient_decode_buf + ((size_t)(cinfo.output_scanline + i) * row_stride);
+        }
+
+        jpeg_read_scanlines(&cinfo, rows, batch);
     }
 
     int out_width = (int)cinfo.output_width;
@@ -922,6 +1358,16 @@ static int ip_jpeg_decode_to_pixels(ip_context_t *ctx, unsigned char **pixels, i
     jpeg_destroy_decompress(&cinfo);
     ctx->jmp_armed = 0;
 
+    unsigned char *buf = ctx->transient_decode_buf;
+    ctx->transient_decode_buf = NULL;
+
+    if (ctx->strip_metadata && ctx->source_orientation > 1) {
+        if (!ip_transform_pixels_for_orientation(ctx, &buf, &out_width, &out_height, ch)) {
+            free(buf);
+            return 0;
+        }
+    }
+
     *pixels = buf;
     *width = out_width;
     *height = out_height;
@@ -965,29 +1411,154 @@ static void ip_clear_output_buffer(ip_context_t *ctx) {
     ctx->output_owner = IP_OUTPUT_OWNER_NONE;
 }
 
-static int ip_decode_jpeg_buffer_preserving_context(ip_context_t *ctx, const unsigned char *data,
-                                                    size_t size, unsigned char **pixels, int *width,
-                                                    int *height, int *channels) {
-    const unsigned char *old_input_data = ctx->input_data;
-    size_t old_input_size = ctx->input_size;
+static int ip_decode_jpeg_to_luma_buffer(ip_context_t *ctx, const unsigned char *data, size_t size,
+                                         unsigned char **luma, int *width, int *height) {
+    struct jpeg_decompress_struct cinfo;
+    ip_jpeg_error_mgr jerr;
+    memset(&cinfo, 0, sizeof(cinfo));
+    memset(&jerr, 0, sizeof(jerr));
+
+    cinfo.err = jpeg_std_error(&jerr.pub);
+    jerr.pub.error_exit = ip_jpeg_invalid_error_exit;
+    jerr.ctx = ctx;
+    ctx->transient_decode_buf = NULL;
+
+    ctx->jmp_armed = 1;
+    if (setjmp(ctx->jmpbuf)) {
+        ctx->jmp_armed = 0;
+        jpeg_destroy_decompress(&cinfo);
+        free(ctx->transient_decode_buf);
+        ctx->transient_decode_buf = NULL;
+        if (ctx->status == IP_OK)
+            ip_context_set_error(ctx, IP_ERR_INVALID_IMAGE, "JPEG luma decode failed");
+        return 0;
+    }
+
+    jpeg_create_decompress(&cinfo);
+    jpeg_mem_src(&cinfo, data, (unsigned long)size);
+
+    int rc = jpeg_read_header(&cinfo, TRUE);
+    if (rc != JPEG_HEADER_OK) {
+        jpeg_destroy_decompress(&cinfo);
+        ctx->jmp_armed = 0;
+        ip_context_set_error(ctx, IP_ERR_INVALID_IMAGE, "invalid JPEG header");
+        return 0;
+    }
+
+    if (cinfo.image_width > (JDIMENSION)INT_MAX || cinfo.image_height > (JDIMENSION)INT_MAX) {
+        jpeg_destroy_decompress(&cinfo);
+        ctx->jmp_armed = 0;
+        ip_context_set_error(ctx, IP_ERR_LIMIT, "JPEG dimensions exceed native int range");
+        return 0;
+    }
+
+    if (cinfo.num_components == 4 || cinfo.jpeg_color_space == JCS_CMYK ||
+        cinfo.jpeg_color_space == JCS_YCCK) {
+        jpeg_destroy_decompress(&cinfo);
+        ctx->jmp_armed = 0;
+        ip_context_set_error(ctx, IP_ERR_UNSUPPORTED,
+                             "CMYK/YCCK JPEG input is not supported in this release");
+        return 0;
+    }
+
     int old_width = ctx->width;
     int old_height = ctx->height;
     int old_channels = ctx->channels;
-    int old_bit_depth = ctx->bit_depth;
     size_t old_decoded_bytes = ctx->decoded_bytes;
 
-    ctx->input_data = data;
-    ctx->input_size = size;
-    int ok = ip_jpeg_decode_to_pixels(ctx, pixels, width, height, channels, 0);
-
-    ctx->input_data = old_input_data;
-    ctx->input_size = old_input_size;
+    ctx->width = (int)cinfo.image_width;
+    ctx->height = (int)cinfo.image_height;
+    ctx->channels = 1;
+    if (!ip_checked_image_size(ctx->width, ctx->height, 1, &ctx->decoded_bytes)) {
+        ctx->width = old_width;
+        ctx->height = old_height;
+        ctx->channels = old_channels;
+        ctx->decoded_bytes = old_decoded_bytes;
+        jpeg_destroy_decompress(&cinfo);
+        ctx->jmp_armed = 0;
+        ip_context_set_error(ctx, IP_ERR_LIMIT, "decoded luma buffer size overflows native size");
+        return 0;
+    }
+    validate_limits_for_pixels(ctx);
     ctx->width = old_width;
     ctx->height = old_height;
     ctx->channels = old_channels;
-    ctx->bit_depth = old_bit_depth;
+    size_t luma_size = ctx->decoded_bytes;
     ctx->decoded_bytes = old_decoded_bytes;
-    return ok;
+
+    if (ctx->status != IP_OK) {
+        jpeg_destroy_decompress(&cinfo);
+        ctx->jmp_armed = 0;
+        return 0;
+    }
+
+    cinfo.out_color_space = JCS_GRAYSCALE;
+#if defined(IMAGE_PACK_HAS_SIMD)
+    cinfo.dct_method = JDCT_ISLOW;
+#else
+    cinfo.dct_method = JDCT_FASTEST;
+#endif
+    cinfo.do_fancy_upsampling = FALSE;
+    cinfo.do_block_smoothing = FALSE;
+    cinfo.quantize_colors = FALSE;
+    cinfo.two_pass_quantize = FALSE;
+    cinfo.dither_mode = JDITHER_NONE;
+
+    jpeg_start_decompress(&cinfo);
+
+    size_t luma_stride = (size_t)cinfo.output_width;
+    if (cinfo.output_components != 1 || luma_stride == 0 ||
+        luma_size != luma_stride * (size_t)cinfo.output_height) {
+        jpeg_destroy_decompress(&cinfo);
+        ctx->jmp_armed = 0;
+        ip_context_set_error(ctx, IP_ERR_LIMIT, "decoded luma buffer size mismatch");
+        return 0;
+    }
+
+    ctx->transient_decode_buf = (unsigned char *)ip_malloc_hot(luma_size);
+    if (!ctx->transient_decode_buf && luma_size > 0) {
+        jpeg_destroy_decompress(&cinfo);
+        ctx->jmp_armed = 0;
+        ip_context_set_error(ctx, IP_ERR_OOM, "failed to allocate luma buffer");
+        return 0;
+    }
+
+    while (cinfo.output_scanline < cinfo.output_height) {
+        if (ctx->cancellable_requested && atomic_load(&ctx->cancelled)) {
+            ip_context_set_error(ctx, IP_ERR_CANCELLED, "JPEG luma decode cancelled");
+            jpeg_abort_decompress(&cinfo);
+            jpeg_destroy_decompress(&cinfo);
+            free(ctx->transient_decode_buf);
+            ctx->transient_decode_buf = NULL;
+            ctx->jmp_armed = 0;
+            return 0;
+        }
+
+        JSAMPROW rows[16];
+        JDIMENSION batch = cinfo.output_height - cinfo.output_scanline;
+        if (batch > 16)
+            batch = 16;
+
+        for (JDIMENSION i = 0; i < batch; i++) {
+            rows[i] =
+                ctx->transient_decode_buf + ((size_t)(cinfo.output_scanline + i) * luma_stride);
+        }
+
+        jpeg_read_scanlines(&cinfo, rows, batch);
+    }
+
+    int out_width = (int)cinfo.output_width;
+    int out_height = (int)cinfo.output_height;
+
+    jpeg_finish_decompress(&cinfo);
+    jpeg_destroy_decompress(&cinfo);
+    ctx->jmp_armed = 0;
+
+    *luma = ctx->transient_decode_buf;
+    ctx->transient_decode_buf = NULL;
+    *width = out_width;
+    *height = out_height;
+    return 1;
 }
 
 static unsigned char *ip_build_luma_buffer(ip_context_t *ctx, const unsigned char *pixels,
@@ -998,7 +1569,7 @@ static unsigned char *ip_build_luma_buffer(ip_context_t *ctx, const unsigned cha
         return NULL;
     }
 
-    unsigned char *luma = (unsigned char *)malloc(count);
+    unsigned char *luma = (unsigned char *)ip_malloc_hot(count);
     if (!luma) {
         ip_context_set_error(ctx, IP_ERR_OOM, "failed to allocate luma buffer");
         return NULL;
@@ -1009,85 +1580,153 @@ static unsigned char *ip_build_luma_buffer(ip_context_t *ctx, const unsigned cha
         return luma;
     }
 
-    for (size_t i = 0; i < count; i++) {
-        const unsigned char *p = pixels + (i * (size_t)channels);
-        unsigned int y =
-            77u * (unsigned int)p[0] + 150u * (unsigned int)p[1] + 29u * (unsigned int)p[2] + 128u;
-        luma[i] = (unsigned char)(y >> 8);
+    const unsigned char *IP_RESTRICT src = pixels;
+    unsigned char *IP_RESTRICT dst = luma;
+
+    if (channels == 3) {
+        for (size_t i = 0; i < count; i++) {
+            unsigned int r = src[i * 3 + 0];
+            unsigned int g = src[i * 3 + 1];
+            unsigned int b = src[i * 3 + 2];
+            dst[i] = (unsigned char)((77u * r + 150u * g + 29u * b + 128u) >> 8);
+        }
+        return luma;
     }
 
+    for (size_t i = 0; i < count; i++) {
+        unsigned int r = src[i * 4 + 0];
+        unsigned int g = src[i * 4 + 1];
+        unsigned int b = src[i * 4 + 2];
+        dst[i] = (unsigned char)((77u * r + 150u * g + 29u * b + 128u) >> 8);
+    }
     return luma;
 }
 
-static double ip_compute_ssim_luma_buffer(const unsigned char *a, const unsigned char *b, int width,
-                                          int height) {
-    const int window = 8;
+static double ip_ssim_window_score_double(int32_t n, int32_t sum_a, int32_t sum_b, int32_t sum_a2,
+                                          int32_t sum_b2, int32_t sum_ab) {
     const double c1 = 6.5025;  /* (0.01 * 255)^2 */
     const double c2 = 58.5225; /* (0.03 * 255)^2 */
-    double total_ssim = 0.0;
-    int windows = 0;
 
-    for (int y0 = 0; y0 < height; y0 += window) {
-        int y1 = y0 + window;
-        if (y1 > height)
-            y1 = height;
+    double inv_n = 1.0 / (double)n;
+    double mean_a = (double)sum_a * inv_n;
+    double mean_b = (double)sum_b * inv_n;
+    double var_a = ((double)sum_a2 * inv_n) - (mean_a * mean_a);
+    double var_b = ((double)sum_b2 * inv_n) - (mean_b * mean_b);
+    double cov_ab = ((double)sum_ab * inv_n) - (mean_a * mean_b);
+
+    if (var_a < 0.0)
+        var_a = 0.0;
+    if (var_b < 0.0)
+        var_b = 0.0;
+
+    double numerator = (2.0 * mean_a * mean_b + c1) * (2.0 * cov_ab + c2);
+    double denominator = (mean_a * mean_a + mean_b * mean_b + c1) * (var_a + var_b + c2);
+    double ssim = denominator == 0.0 ? 1.0 : numerator / denominator;
+
+    if (ssim < 0.0)
+        ssim = 0.0;
+    if (ssim > 1.0)
+        ssim = 1.0;
+    return ssim;
+}
 
-        for (int x0 = 0; x0 < width; x0 += window) {
-            int x1 = x0 + window;
-            if (x1 > width)
-                x1 = width;
-
-            double sum_a = 0.0;
-            double sum_b = 0.0;
-            double sum_a2 = 0.0;
-            double sum_b2 = 0.0;
-            double sum_ab = 0.0;
-            int n = 0;
-
-            for (int y = y0; y < y1; y++) {
-                size_t row = (size_t)y * (size_t)width;
-                for (int x = x0; x < x1; x++) {
-                    size_t idx = row + (size_t)x;
-                    double la = (double)a[idx];
-                    double lb = (double)b[idx];
-                    sum_a += la;
-                    sum_b += lb;
-                    sum_a2 += la * la;
-                    sum_b2 += lb * lb;
-                    sum_ab += la * lb;
-                    n++;
-                }
-            }
+static inline double ip_ssim_window_8x8(const unsigned char *IP_RESTRICT a,
+                                        const unsigned char *IP_RESTRICT b, int width, int x0,
+                                        int y0) {
+    int32_t sum_a = 0, sum_b = 0, sum_a2 = 0, sum_b2 = 0, sum_ab = 0;
+
+    for (int y = 0; y < 8; y++) {
+        const unsigned char *pa = a + (size_t)(y0 + y) * (size_t)width + (size_t)x0;
+        const unsigned char *pb = b + (size_t)(y0 + y) * (size_t)width + (size_t)x0;
+        for (int x = 0; x < 8; x++) {
+            int32_t la = pa[x];
+            int32_t lb = pb[x];
+            sum_a += la;
+            sum_b += lb;
+            sum_a2 += la * la;
+            sum_b2 += lb * lb;
+            sum_ab += la * lb;
+        }
+    }
 
-            if (n <= 0)
-                continue;
+    return ip_ssim_window_score_double(64, sum_a, sum_b, sum_a2, sum_b2, sum_ab);
+}
 
-            double inv_n = 1.0 / (double)n;
-            double mean_a = sum_a * inv_n;
-            double mean_b = sum_b * inv_n;
-            double var_a = (sum_a2 * inv_n) - (mean_a * mean_a);
-            double var_b = (sum_b2 * inv_n) - (mean_b * mean_b);
-            double cov_ab = (sum_ab * inv_n) - (mean_a * mean_b);
+static inline double ip_ssim_window_var(const unsigned char *IP_RESTRICT a,
+                                        const unsigned char *IP_RESTRICT b, int width, int x0,
+                                        int y0, int x1, int y1) {
+    int32_t sum_a = 0, sum_b = 0, sum_a2 = 0, sum_b2 = 0, sum_ab = 0;
+    int32_t n = 0;
+
+    for (int y = y0; y < y1; y++) {
+        const unsigned char *pa = a + (size_t)y * (size_t)width;
+        const unsigned char *pb = b + (size_t)y * (size_t)width;
+        for (int x = x0; x < x1; x++) {
+            int32_t la = pa[x];
+            int32_t lb = pb[x];
+            sum_a += la;
+            sum_b += lb;
+            sum_a2 += la * la;
+            sum_b2 += lb * lb;
+            sum_ab += la * lb;
+            n++;
+        }
+    }
 
-            if (var_a < 0.0)
-                var_a = 0.0;
-            if (var_b < 0.0)
-                var_b = 0.0;
+    if (n <= 0)
+        return 1.0;
+    return ip_ssim_window_score_double(n, sum_a, sum_b, sum_a2, sum_b2, sum_ab);
+}
+
+static double ip_compute_ssim_luma_buffer(const unsigned char *a, const unsigned char *b, int width,
+                                          int height) {
+    const int window = 8;
+    double total_ssim = 0.0;
+    int windows = 0;
 
-            double numerator = (2.0 * mean_a * mean_b + c1) * (2.0 * cov_ab + c2);
-            double denominator = (mean_a * mean_a + mean_b * mean_b + c1) * (var_a + var_b + c2);
-            double ssim = denominator == 0.0 ? 1.0 : numerator / denominator;
+    int full_x = width / window;
+    int full_y = height / window;
+    int rem_x = width - full_x * window;
+    int rem_y = height - full_y * window;
 
-            if (ssim < 0.0)
-                ssim = 0.0;
-            if (ssim > 1.0)
-                ssim = 1.0;
+    for (int by = 0; by < full_y; by++) {
+        int y0 = by * window;
+        for (int bx = 0; bx < full_x; bx++) {
+            int x0 = bx * window;
+            total_ssim += ip_ssim_window_8x8(a, b, width, x0, y0);
+            windows++;
+        }
+    }
+
+    if (rem_x > 0) {
+        int x0 = full_x * window;
+        int x1 = width;
+        for (int by = 0; by < full_y; by++) {
+            int y0 = by * window;
+            int y1 = y0 + window;
+            total_ssim += ip_ssim_window_var(a, b, width, x0, y0, x1, y1);
+            windows++;
+        }
+    }
 
-            total_ssim += ssim;
+    if (rem_y > 0) {
+        int y0 = full_y * window;
+        int y1 = height;
+        for (int bx = 0; bx < full_x; bx++) {
+            int x0 = bx * window;
+            int x1 = x0 + window;
+            total_ssim += ip_ssim_window_var(a, b, width, x0, y0, x1, y1);
             windows++;
         }
     }
 
+    if (rem_x > 0 && rem_y > 0) {
+        int x0 = full_x * window;
+        int y0 = full_y * window;
+        total_ssim += ip_ssim_window_var(a, b, width, x0, y0, width, height);
+        windows++;
+    }
+
     return windows > 0 ? total_ssim / (double)windows : 0.0;
 }
 
@@ -1098,13 +1737,13 @@ static int guarded_compress_jpeg_input_with_mode(ip_context_t *ctx, int mozjpeg_
     int reference_channels = 0;
 
     if (ctx->pixel_data) {
-        reference_pixels = ctx->owned_pixel_data;
+        reference_pixels = (unsigned char *)ctx->pixel_data;
         reference_width = ctx->width;
         reference_height = ctx->height;
         reference_channels = ctx->channels;
     } else {
         if (!ip_jpeg_decode_to_pixels(ctx, &reference_pixels, &reference_width, &reference_height,
-                                      &reference_channels, 0)) {
+                                      &reference_channels, 1)) {
             return 0;
         }
 
@@ -1120,7 +1759,7 @@ static int guarded_compress_jpeg_input_with_mode(ip_context_t *ctx, int mozjpeg_
 
     if (reference_channels == 4) {
         ip_context_set_error(ctx, IP_ERR_UNSUPPORTED,
-                             "min_ssim is not supported for RGBA input in v0.2.0");
+                             "min_ssim is not supported for RGBA input in v0.2.2");
         return 0;
     }
 
@@ -1163,13 +1802,12 @@ static int guarded_compress_jpeg_input_with_mode(ip_context_t *ctx, int mozjpeg_
         ctx->output_capacity = 0;
         ctx->output_owner = IP_OUTPUT_OWNER_NONE;
 
-        unsigned char *candidate_pixels = NULL;
+        unsigned char *candidate_luma = NULL;
         int candidate_width = 0;
         int candidate_height = 0;
-        int candidate_channels = 0;
-        int decoded_ok = ip_decode_jpeg_buffer_preserving_context(
-            ctx, candidate_jpeg, candidate_jpeg_size, &candidate_pixels, &candidate_width,
-            &candidate_height, &candidate_channels);
+        int decoded_ok =
+            ip_decode_jpeg_to_luma_buffer(ctx, candidate_jpeg, candidate_jpeg_size, &candidate_luma,
+                                          &candidate_width, &candidate_height);
 
         if (!decoded_ok) {
             free(reference_luma);
@@ -1178,10 +1816,9 @@ static int guarded_compress_jpeg_input_with_mode(ip_context_t *ctx, int mozjpeg_
             return 0;
         }
 
-        if (candidate_width != reference_width || candidate_height != reference_height ||
-            candidate_channels != reference_channels) {
+        if (candidate_width != reference_width || candidate_height != reference_height) {
             free(reference_luma);
-            free(candidate_pixels);
+            free(candidate_luma);
             free(candidate_jpeg);
             free(best_jpeg);
             ip_context_set_error(ctx, IP_ERR_ENCODE,
@@ -1189,17 +1826,6 @@ static int guarded_compress_jpeg_input_with_mode(ip_context_t *ctx, int mozjpeg_
             return 0;
         }
 
-        unsigned char *candidate_luma = ip_build_luma_buffer(ctx, candidate_pixels, candidate_width,
-                                                             candidate_height, candidate_channels);
-        free(candidate_pixels);
-
-        if (!candidate_luma) {
-            free(reference_luma);
-            free(candidate_jpeg);
-            free(best_jpeg);
-            return 0;
-        }
-
         double ssim = ip_compute_ssim_luma_buffer(reference_luma, candidate_luma, reference_width,
                                                   reference_height);
         free(candidate_luma);
@@ -1244,119 +1870,176 @@ static int guarded_compress_jpeg_input_with_mode(ip_context_t *ctx, int mozjpeg_
     return 1;
 }
 
-#if defined(IMAGE_PACK_ENABLE_LOSSLESS_TRANSCODE_FAST_PATH)
-static int ip_jpeg_transcode_coefficients(ip_context_t *ctx, int mozjpeg_size_mode) {
-    if (mozjpeg_size_mode)
+static void ip_setup_marker_saving(struct jpeg_decompress_struct *cinfo, int strip_metadata) {
+    jpeg_save_markers(cinfo, JPEG_APP0 + 1, 0xFFFF);
+    if (!strip_metadata) {
+        jpeg_save_markers(cinfo, JPEG_COM, 0xFFFF);
+        for (int app = 2; app < 16; app++) {
+            jpeg_save_markers(cinfo, JPEG_APP0 + app, 0xFFFF);
+        }
+    }
+}
+
+static int ip_validate_lossless_optimize_header(ip_context_t *ctx,
+                                                struct jpeg_decompress_struct *srcinfo) {
+    if (srcinfo->image_width > (JDIMENSION)INT_MAX || srcinfo->image_height > (JDIMENSION)INT_MAX) {
+        ip_context_set_error(ctx, IP_ERR_LIMIT, "JPEG dimensions exceed native int range");
         return 0;
-    if (!ctx->input_data || ctx->input_size < 4)
+    }
+
+    ctx->width = (int)srcinfo->image_width;
+    ctx->height = (int)srcinfo->image_height;
+    ctx->channels = srcinfo->num_components;
+    ctx->bit_depth = 8;
+
+    if (ctx->max_width < 0 || ctx->max_height < 0) {
+        ip_context_set_error(ctx, IP_ERR_INVALID_ARGUMENT, "max_width/max_height must be >= 0");
         return 0;
-    if (ctx->input_data[0] != 0xFF || ctx->input_data[1] != 0xD8)
+    }
+    if (ctx->max_width > 0 && ctx->width > ctx->max_width) {
+        ip_context_set_error(ctx, IP_ERR_LIMIT, "image width exceeds max_width");
+        return 0;
+    }
+    if (ctx->max_height > 0 && ctx->height > ctx->max_height) {
+        ip_context_set_error(ctx, IP_ERR_LIMIT, "image height exceeds max_height");
         return 0;
+    }
+    if (ctx->max_pixels > 0 && (uint64_t)ctx->width * (uint64_t)ctx->height > ctx->max_pixels) {
+        ip_context_set_error(ctx, IP_ERR_LIMIT, "image pixels exceed max_pixels");
+        return 0;
+    }
 
+    return 1;
+}
+
+static int ip_lossless_optimize_jpeg(ip_context_t *ctx) {
     struct jpeg_decompress_struct srcinfo;
     struct jpeg_compress_struct dstinfo;
-    ip_jpeg_error_mgr jsrcerr;
-    ip_jpeg_error_mgr jdsterr;
+    ip_jpeg_error_mgr srcerr;
+    ip_jpeg_error_mgr dsterr;
+    jvirt_barray_ptr *coef_arrays = NULL;
+    unsigned long jpeg_size = 0;
+
     memset(&srcinfo, 0, sizeof(srcinfo));
     memset(&dstinfo, 0, sizeof(dstinfo));
-    memset(&jsrcerr, 0, sizeof(jsrcerr));
-    memset(&jdsterr, 0, sizeof(jdsterr));
+    memset(&srcerr, 0, sizeof(srcerr));
+    memset(&dsterr, 0, sizeof(dsterr));
+    ctx->transient_jpeg_buf = NULL;
 
-    srcinfo.err = jpeg_std_error(&jsrcerr.pub);
-    jsrcerr.pub.error_exit = ip_jpeg_invalid_error_exit;
-    jsrcerr.ctx = ctx;
+    srcinfo.err = jpeg_std_error(&srcerr.pub);
+    srcerr.pub.error_exit = ip_jpeg_invalid_error_exit;
+    srcerr.ctx = ctx;
 
-    dstinfo.err = jpeg_std_error(&jdsterr.pub);
-    jdsterr.pub.error_exit = ip_jpeg_encode_error_exit;
-    jdsterr.ctx = ctx;
-
-    unsigned char *jpeg_buf = NULL;
-    unsigned long jpeg_size = 0;
-    int src_created = 0;
-    int dst_created = 0;
+    dstinfo.err = jpeg_std_error(&dsterr.pub);
+    dsterr.pub.error_exit = ip_jpeg_encode_error_exit;
+    dsterr.ctx = ctx;
 
     ctx->jmp_armed = 1;
     if (setjmp(ctx->jmpbuf)) {
         ctx->jmp_armed = 0;
-        if (dst_created)
-            jpeg_destroy_compress(&dstinfo);
-        if (src_created)
-            jpeg_destroy_decompress(&srcinfo);
-        free(jpeg_buf);
-
+        jpeg_destroy_compress(&dstinfo);
+        jpeg_destroy_decompress(&srcinfo);
+        free(ctx->transient_jpeg_buf);
+        ctx->transient_jpeg_buf = NULL;
         if (ctx->status == IP_OK)
-            return 0;
+            ip_context_set_error(ctx, IP_ERR_ENCODE, "lossless JPEG optimize failed");
         return 0;
     }
 
     jpeg_create_decompress(&srcinfo);
-    src_created = 1;
+    jpeg_create_compress(&dstinfo);
     jpeg_mem_src(&srcinfo, ctx->input_data, (unsigned long)ctx->input_size);
+    ip_setup_marker_saving(&srcinfo, ctx->strip_metadata);
 
-    if (jpeg_read_header(&srcinfo, TRUE) != JPEG_HEADER_OK) {
+    if (ctx->cancellable_requested && atomic_load(&ctx->cancelled)) {
+        ip_context_set_error(ctx, IP_ERR_CANCELLED, "lossless JPEG optimize cancelled");
+        jpeg_destroy_compress(&dstinfo);
         jpeg_destroy_decompress(&srcinfo);
         ctx->jmp_armed = 0;
         return 0;
     }
 
-    if ((int)srcinfo.image_width > ctx->max_width || (int)srcinfo.image_height > ctx->max_height ||
-        (uint64_t)srcinfo.image_width * (uint64_t)srcinfo.image_height > ctx->max_pixels) {
+    int rc = jpeg_read_header(&srcinfo, TRUE);
+    if (rc != JPEG_HEADER_OK) {
+        jpeg_destroy_compress(&dstinfo);
         jpeg_destroy_decompress(&srcinfo);
         ctx->jmp_armed = 0;
-        ip_context_set_error(ctx, IP_ERR_LIMIT, "image exceeds configured limits");
+        ip_context_set_error(ctx, IP_ERR_INVALID_IMAGE, "invalid JPEG header");
         return 0;
     }
 
-    jvirt_barray_ptr *src_coef_arrays = jpeg_read_coefficients(&srcinfo);
-    if (!src_coef_arrays) {
+    ctx->source_orientation = ip_read_exif_orientation_from_decompress(&srcinfo);
+    if (ctx->strip_metadata && ctx->source_orientation > 1) {
+        jpeg_destroy_compress(&dstinfo);
         jpeg_destroy_decompress(&srcinfo);
         ctx->jmp_armed = 0;
+        ip_context_set_error(ctx, IP_ERR_UNSUPPORTED,
+                             "lossless optimize cannot strip EXIF Orientation without changing "
+                             "visual orientation; use strip_metadata: false or ImagePack.compress");
         return 0;
     }
 
-    jpeg_create_compress(&dstinfo);
-    dst_created = 1;
-    jpeg_mem_dest(&dstinfo, &jpeg_buf, &jpeg_size);
-    jpeg_copy_critical_parameters(&srcinfo, &dstinfo);
+    if (!ip_validate_lossless_optimize_header(ctx, &srcinfo)) {
+        jpeg_destroy_compress(&dstinfo);
+        jpeg_destroy_decompress(&srcinfo);
+        ctx->jmp_armed = 0;
+        return 0;
+    }
+
+    if (!ctx->strip_metadata && !ip_save_markers_from_decompress(ctx, &srcinfo)) {
+        jpeg_destroy_compress(&dstinfo);
+        jpeg_destroy_decompress(&srcinfo);
+        ctx->jmp_armed = 0;
+        return 0;
+    }
 
-    dstinfo.optimize_coding = FALSE;
+    coef_arrays = jpeg_read_coefficients(&srcinfo);
+    jpeg_copy_critical_parameters(&srcinfo, &dstinfo);
+    dstinfo.optimize_coding = TRUE;
+    dstinfo.num_scans = 0;
+    dstinfo.scan_info = NULL;
     if (ctx->progressive) {
         jpeg_simple_progression(&dstinfo);
         dstinfo.optimize_coding = TRUE;
     }
 
-    jpeg_write_coefficients(&dstinfo, src_coef_arrays);
-    jpeg_finish_compress(&dstinfo);
-    jpeg_destroy_compress(&dstinfo);
-    dst_created = 0;
+    jpeg_mem_dest(&dstinfo, &ctx->transient_jpeg_buf, &jpeg_size);
+
+    if (ctx->cancellable_requested && atomic_load(&ctx->cancelled)) {
+        ip_context_set_error(ctx, IP_ERR_CANCELLED, "lossless JPEG optimize cancelled");
+        jpeg_destroy_compress(&dstinfo);
+        jpeg_destroy_decompress(&srcinfo);
+        free(ctx->transient_jpeg_buf);
+        ctx->transient_jpeg_buf = NULL;
+        ctx->jmp_armed = 0;
+        return 0;
+    }
+
+    jpeg_write_coefficients(&dstinfo, coef_arrays);
+    if (!ctx->strip_metadata)
+        ip_write_preserved_markers(ctx, &dstinfo);
 
+    jpeg_finish_compress(&dstinfo);
     jpeg_finish_decompress(&srcinfo);
+    jpeg_destroy_compress(&dstinfo);
     jpeg_destroy_decompress(&srcinfo);
-    src_created = 0;
-
     ctx->jmp_armed = 0;
 
-    if ((size_t)jpeg_size > ctx->max_output_size) {
-        free(jpeg_buf);
-        ip_context_set_error(ctx, IP_ERR_LIMIT, "transcoded output exceeds max_output_size");
+    if (ctx->max_output_size > 0 && (size_t)jpeg_size > ctx->max_output_size) {
+        free(ctx->transient_jpeg_buf);
+        ctx->transient_jpeg_buf = NULL;
+        ip_context_set_error(ctx, IP_ERR_LIMIT, "output exceeds max_output_size");
         return 0;
     }
 
-    ctx->output_data = jpeg_buf;
+    ctx->output_data = ctx->transient_jpeg_buf;
+    ctx->transient_jpeg_buf = NULL;
     ctx->output_size = (size_t)jpeg_size;
     ctx->output_capacity = (size_t)jpeg_size;
     ctx->output_owner = IP_OUTPUT_OWNER_MALLOC;
-
-    if (ctx->width == 0 || ctx->height == 0) {
-        ctx->width = 0;
-        ctx->height = 0;
-        ctx->channels = 0;
-    }
     return 1;
 }
 
-#endif
-
 static int ip_jpeg_turbo_compress(ip_context_t *ctx) {
     if (ctx->ssim_guard_enabled)
         return guarded_compress_jpeg_input_with_mode(ctx, 0);
@@ -1436,6 +2119,29 @@ static int ip_run_context(ip_context_t *ctx) {
     return ctx->status == IP_OK;
 }
 
+static void *ip_run_optimize_nogvl(void *data) {
+    ip_context_t *ctx = (ip_context_t *)data;
+    if (ctx->status == IP_OK)
+        ip_lossless_optimize_jpeg(ctx);
+    return NULL;
+}
+
+static int ip_run_optimize_context(ip_context_t *ctx) {
+    ip_resolve_execution(ctx);
+
+    if (ctx->resolved_execution == IP_EXEC_DIRECT) {
+        ip_run_optimize_nogvl(ctx);
+    } else if (ctx->resolved_execution == IP_EXEC_NOGVL) {
+        rb_nogvl(ip_run_optimize_nogvl, ctx, ip_unblock_function, ctx, 0);
+    } else if (ctx->resolved_execution == IP_EXEC_OFFLOAD) {
+        rb_nogvl(ip_run_optimize_nogvl, ctx, ip_unblock_function, ctx, RB_NOGVL_OFFLOAD_SAFE);
+    } else {
+        ip_context_set_error(ctx, IP_ERR_INVALID_ARGUMENT, "invalid resolved execution mode");
+    }
+
+    return ctx->status == IP_OK;
+}
+
 static size_t config_size_value(VALUE config, ID id, size_t fallback) {
     VALUE value = rb_funcall(config, id, 0);
     if (NIL_P(value))
@@ -1460,6 +2166,7 @@ static void apply_configuration(VALUE self, ip_context_t *ctx) {
     ctx->max_width = config_int_value(config, id_max_width, ctx->max_width);
     ctx->max_height = config_int_value(config, id_max_height, ctx->max_height);
     ctx->max_output_size = config_size_value(config, id_max_output_size, ctx->max_output_size);
+    ctx->max_input_size = config_size_value(config, id_max_input_size, ctx->max_input_size);
 }
 
 static void validate_limits_for_pixels(ip_context_t *ctx) {
@@ -1469,15 +2176,19 @@ static void validate_limits_for_pixels(ip_context_t *ctx) {
         ip_context_set_error(ctx, IP_ERR_INVALID_ARGUMENT, "image dimensions must be positive");
         return;
     }
-    if (ctx->width > ctx->max_width) {
+    if (ctx->max_width < 0 || ctx->max_height < 0) {
+        ip_context_set_error(ctx, IP_ERR_INVALID_ARGUMENT, "max_width/max_height must be >= 0");
+        return;
+    }
+    if (ctx->max_width > 0 && ctx->width > ctx->max_width) {
         ip_context_set_error(ctx, IP_ERR_LIMIT, "image width exceeds max_width");
         return;
     }
-    if (ctx->height > ctx->max_height) {
+    if (ctx->max_height > 0 && ctx->height > ctx->max_height) {
         ip_context_set_error(ctx, IP_ERR_LIMIT, "image height exceeds max_height");
         return;
     }
-    if ((uint64_t)ctx->width * (uint64_t)ctx->height > ctx->max_pixels) {
+    if (ctx->max_pixels > 0 && (uint64_t)ctx->width * (uint64_t)ctx->height > ctx->max_pixels) {
         ip_context_set_error(ctx, IP_ERR_LIMIT, "image pixels exceed max_pixels");
         return;
     }
@@ -1489,95 +2200,137 @@ static void validate_limits_for_pixels(ip_context_t *ctx) {
     ctx->decoded_bytes = decoded_bytes;
 }
 
-static VALUE ip_compress_jpeg_entry(VALUE self, VALUE input, VALUE input_kind, VALUE output,
-                                    VALUE output_kind, VALUE algo, VALUE quality, VALUE min_ssim,
-                                    VALUE mozjpeg_trellis, VALUE progressive, VALUE strip_metadata,
-                                    VALUE execution, VALUE cancellable, VALUE has_scheduler) {
+static VALUE ip_compress_jpeg_entry_body(VALUE ptr) {
+    ip_compress_jpeg_call_t *call = (ip_compress_jpeg_call_t *)ptr;
     ip_context_t *ctx = ip_context_new();
     if (!ctx)
         rb_raise(rb_eImagePackOutOfMemoryError, "failed to allocate native context");
+    call->ctx = ctx;
 
-    ip_output_kind_t out_kind = ip_parse_output_kind(output_kind);
-    ctx->algo = ip_parse_algo(algo);
-    ctx->quality = NUM2INT(quality);
+    ip_output_kind_t out_kind = ip_parse_output_kind(call->output_kind);
+    ctx->algo = ip_parse_algo(call->algo);
+    ctx->quality = NUM2INT(call->quality);
     ctx->selected_quality = ctx->quality;
     ip_validate_quality_or_raise(ctx);
-    ctx->min_ssim = NUM2DBL(min_ssim);
+    ctx->min_ssim = NUM2DBL(call->min_ssim);
     ctx->ssim_guard_enabled = ctx->min_ssim > 0.0;
     ip_validate_min_ssim_or_raise(ctx);
-    ctx->mozjpeg_trellis_enabled = ip_bool_value(mozjpeg_trellis);
-    ctx->progressive = ip_bool_value(progressive);
-    ctx->strip_metadata = ip_bool_value(strip_metadata);
-    ctx->requested_execution = ip_parse_execution(execution);
-    ctx->cancellable_requested = ip_bool_value(cancellable);
-    ctx->has_scheduler = ip_bool_value(has_scheduler);
-    apply_configuration(self, ctx);
-
-    if (!ip_prepare_input_bytes(ctx, input, ip_parse_input_kind(input_kind)) ||
-        !ip_prepare_output_path(ctx, output, out_kind)) {
-        VALUE exception = ip_status_to_exception(ctx->status);
-        char message[512];
-        snprintf(message, sizeof(message), "%s", ctx->error_message);
-        ip_context_free(ctx);
-        rb_raise(exception, "%s", message[0] ? message : "invalid JPEG input");
+    ctx->mozjpeg_trellis_enabled = ip_bool_value(call->mozjpeg_trellis);
+    ctx->progressive = ip_bool_value(call->progressive);
+    ctx->strip_metadata = ip_bool_value(call->strip_metadata);
+    ctx->requested_execution = ip_parse_execution(call->execution);
+    ctx->cancellable_requested = ip_bool_value(call->cancellable);
+    ctx->has_scheduler = ip_bool_value(call->has_scheduler);
+    apply_configuration(call->self, ctx);
+
+    if (!ip_prepare_input_bytes(ctx, call->input, ip_parse_input_kind(call->input_kind)) ||
+        !ip_prepare_output_path(ctx, call->output, out_kind)) {
+        ip_raise_for_status(ctx);
+        rb_raise(rb_eImagePackInvalidArgumentError, "invalid JPEG input");
     }
 
     if (ctx->requested_execution == IP_EXEC_AUTO && ctx->input_size < ctx->direct_input_threshold &&
         !ip_inspect_jpeg_header(ctx)) {
-        VALUE exception = ip_status_to_exception(ctx->status);
-        char message[512];
-        snprintf(message, sizeof(message), "%s", ctx->error_message);
-        ip_context_free(ctx);
-        rb_raise(exception, "%s", message[0] ? message : "invalid JPEG input");
+        ip_raise_for_status(ctx);
+        rb_raise(rb_eImagePackInvalidImageError, "invalid JPEG input");
     }
 
     ip_run_context(ctx);
-    VALUE result = ip_finish_output(ctx, out_kind);
-    ip_context_free(ctx);
-    return result;
+    return ip_finish_output(ctx, out_kind);
 }
 
-static VALUE ip_compress_pixels_entry(VALUE self, VALUE buffer, VALUE width, VALUE height,
-                                      VALUE channels, VALUE output, VALUE output_kind, VALUE algo,
-                                      VALUE quality, VALUE progressive, VALUE execution,
-                                      VALUE cancellable, VALUE has_scheduler) {
+static VALUE ip_compress_jpeg_entry(VALUE self, VALUE input, VALUE input_kind, VALUE output,
+                                    VALUE output_kind, VALUE algo, VALUE quality, VALUE min_ssim,
+                                    VALUE mozjpeg_trellis, VALUE progressive, VALUE strip_metadata,
+                                    VALUE execution, VALUE cancellable, VALUE has_scheduler) {
+    ip_compress_jpeg_call_t call = {
+        self,           input,     input_kind,  output,          output_kind,
+        algo,           quality,   min_ssim,    mozjpeg_trellis, progressive,
+        strip_metadata, execution, cancellable, has_scheduler,   NULL};
+    return rb_ensure(ip_compress_jpeg_entry_body, (VALUE)&call, ip_call_cleanup, (VALUE)&call.ctx);
+}
+
+static VALUE ip_compress_pixels_entry_body(VALUE ptr) {
+    ip_compress_pixels_call_t *call = (ip_compress_pixels_call_t *)ptr;
     ip_context_t *ctx = ip_context_new();
     if (!ctx)
         rb_raise(rb_eImagePackOutOfMemoryError, "failed to allocate native context");
+    call->ctx = ctx;
 
-    ip_output_kind_t out_kind = ip_parse_output_kind(output_kind);
-    ctx->algo = ip_parse_algo(algo);
-    ctx->quality = NUM2INT(quality);
+    ip_output_kind_t out_kind = ip_parse_output_kind(call->output_kind);
+    ctx->algo = ip_parse_algo(call->algo);
+    ctx->quality = NUM2INT(call->quality);
+    ctx->selected_quality = ctx->quality;
     ip_validate_quality_or_raise(ctx);
-    ctx->progressive = ip_bool_value(progressive);
+    ctx->min_ssim = NUM2DBL(call->min_ssim);
+    ctx->ssim_guard_enabled = ctx->min_ssim > 0.0;
+    ip_validate_min_ssim_or_raise(ctx);
+    ctx->progressive = ip_bool_value(call->progressive);
     ctx->strip_metadata = 1;
-    ctx->requested_execution = ip_parse_execution(execution);
-    ctx->cancellable_requested = ip_bool_value(cancellable);
-    ctx->has_scheduler = ip_bool_value(has_scheduler);
-    apply_configuration(self, ctx);
-
-    if (!ip_prepare_pixels(ctx, buffer, NUM2INT(width), NUM2INT(height), NUM2INT(channels)) ||
-        !ip_prepare_output_path(ctx, output, out_kind)) {
-        VALUE exception = ip_status_to_exception(ctx->status);
-        char message[512];
-        snprintf(message, sizeof(message), "%s", ctx->error_message);
-        ip_context_free(ctx);
-        rb_raise(exception, "%s", message[0] ? message : "invalid pixel input");
+    ctx->requested_execution = ip_parse_execution(call->execution);
+    ctx->cancellable_requested = ip_bool_value(call->cancellable);
+    ctx->has_scheduler = ip_bool_value(call->has_scheduler);
+    apply_configuration(call->self, ctx);
+
+    if (!ip_prepare_pixels(ctx, call->buffer, NUM2INT(call->width), NUM2INT(call->height),
+                           NUM2INT(call->channels)) ||
+        !ip_prepare_output_path(ctx, call->output, out_kind)) {
+        ip_raise_for_status(ctx);
+        rb_raise(rb_eImagePackInvalidArgumentError, "invalid pixel input");
     }
 
     validate_limits_for_pixels(ctx);
-    if (ctx->status != IP_OK) {
-        VALUE exception = ip_status_to_exception(ctx->status);
-        char message[512];
-        snprintf(message, sizeof(message), "%s", ctx->error_message);
-        ip_context_free(ctx);
-        rb_raise(exception, "%s", message);
-    }
+    if (ctx->status != IP_OK)
+        ip_raise_for_status(ctx);
 
     ip_run_context(ctx);
-    VALUE result = ip_finish_output(ctx, out_kind);
-    ip_context_free(ctx);
-    return result;
+    return ip_finish_output(ctx, out_kind);
+}
+
+static VALUE ip_compress_pixels_entry(VALUE self, VALUE buffer, VALUE width, VALUE height,
+                                      VALUE channels, VALUE output, VALUE output_kind, VALUE algo,
+                                      VALUE quality, VALUE min_ssim, VALUE progressive,
+                                      VALUE execution, VALUE cancellable, VALUE has_scheduler) {
+    ip_compress_pixels_call_t call = {
+        self,    buffer,   width,       height,    channels,    output,        output_kind, algo,
+        quality, min_ssim, progressive, execution, cancellable, has_scheduler, NULL};
+    return rb_ensure(ip_compress_pixels_entry_body, (VALUE)&call, ip_call_cleanup,
+                     (VALUE)&call.ctx);
+}
+
+static VALUE ip_optimize_jpeg_entry_body(VALUE ptr) {
+    ip_optimize_jpeg_call_t *call = (ip_optimize_jpeg_call_t *)ptr;
+    ip_context_t *ctx = ip_context_new();
+    if (!ctx)
+        rb_raise(rb_eImagePackOutOfMemoryError, "failed to allocate native context");
+    call->ctx = ctx;
+
+    ip_output_kind_t out_kind = ip_parse_output_kind(call->output_kind);
+    ctx->progressive = ip_bool_value(call->progressive);
+    ctx->strip_metadata = ip_bool_value(call->strip_metadata);
+    ctx->requested_execution = ip_parse_execution(call->execution);
+    ctx->cancellable_requested = ip_bool_value(call->cancellable);
+    ctx->has_scheduler = ip_bool_value(call->has_scheduler);
+    ctx->ssim_guard_enabled = 0;
+    apply_configuration(call->self, ctx);
+
+    if (!ip_prepare_input_bytes(ctx, call->input, ip_parse_input_kind(call->input_kind)) ||
+        !ip_prepare_output_path(ctx, call->output, out_kind)) {
+        ip_raise_for_status(ctx);
+        rb_raise(rb_eImagePackInvalidArgumentError, "invalid JPEG input");
+    }
+
+    ip_run_optimize_context(ctx);
+    return ip_finish_output(ctx, out_kind);
+}
+
+static VALUE ip_optimize_jpeg_entry(VALUE self, VALUE input, VALUE input_kind, VALUE output,
+                                    VALUE output_kind, VALUE progressive, VALUE strip_metadata,
+                                    VALUE execution, VALUE cancellable, VALUE has_scheduler) {
+    ip_optimize_jpeg_call_t call = {
+        self,           input,     input_kind,  output,        output_kind, progressive,
+        strip_metadata, execution, cancellable, has_scheduler, NULL};
+    return rb_ensure(ip_optimize_jpeg_entry_body, (VALUE)&call, ip_call_cleanup, (VALUE)&call.ctx);
 }
 
 IMAGE_PACK_INIT_EXPORT void Init_image_pack(void) {
@@ -1598,6 +2351,7 @@ IMAGE_PACK_INIT_EXPORT void Init_image_pack(void) {
     id_max_width = rb_intern("max_width");
     id_max_height = rb_intern("max_height");
     id_max_output_size = rb_intern("max_output_size");
+    id_max_input_size = rb_intern("max_input_size");
 
     rb_mImagePack = rb_define_module("ImagePack");
     rb_eImagePackError = rb_const_get(rb_mImagePack, rb_intern("Error"));
@@ -1612,7 +2366,23 @@ IMAGE_PACK_INIT_EXPORT void Init_image_pack(void) {
     rb_eImagePackOutOfMemoryError = rb_const_get(rb_mImagePack, rb_intern("OutOfMemoryError"));
     rb_eImagePackCancelledError = rb_const_get(rb_mImagePack, rb_intern("CancelledError"));
 
+    rb_define_const(rb_mImagePack, "NATIVE_MOZJPEG_VERSION", rb_str_new_cstr(VERSION));
+#if defined(IMAGE_PACK_HAS_SIMD)
+    rb_define_const(rb_mImagePack, "NATIVE_SIMD", Qtrue);
+#else
+    rb_define_const(rb_mImagePack, "NATIVE_SIMD", Qfalse);
+#endif
+
     rb_define_singleton_method(rb_mImagePack, "__compress_jpeg", ip_compress_jpeg_entry, 13);
-    rb_define_singleton_method(rb_mImagePack, "__compress_pixels", ip_compress_pixels_entry, 12);
+    rb_define_singleton_method(rb_mImagePack, "__compress_pixels", ip_compress_pixels_entry, 13);
+    rb_define_singleton_method(rb_mImagePack, "__optimize_jpeg", ip_optimize_jpeg_entry, 9);
     rb_define_singleton_method(rb_mImagePack, "__inspect_image", ip_inspect_image_entry, 2);
+    rb_funcall(rb_mImagePack, rb_intern("private_class_method"), 1,
+               ID2SYM(rb_intern("__compress_jpeg")));
+    rb_funcall(rb_mImagePack, rb_intern("private_class_method"), 1,
+               ID2SYM(rb_intern("__compress_pixels")));
+    rb_funcall(rb_mImagePack, rb_intern("private_class_method"), 1,
+               ID2SYM(rb_intern("__optimize_jpeg")));
+    rb_funcall(rb_mImagePack, rb_intern("private_class_method"), 1,
+               ID2SYM(rb_intern("__inspect_image")));
 }