nanoodle 0.7.0 → 0.9.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/src/graph.mjs CHANGED
@@ -3,7 +3,7 @@ import { NanoodleError } from "./errors.mjs";
3
3
  /* Dynamic input-port families (mirrors the nanoodle app's runGraph). A wire landing on one of
4
4
  these ports — or on a port declared in NODE_TYPES[type].inputs — is a data input; a wire
5
5
  landing on ANY other port is a field override (wired prompt/system/lyrics/q/...). */
6
- export const IMG_PORT_RE = /^img\d+$/; // llm vision references
6
+ export const IMG_PORT_RE = /^img\d+$/; // llm vision references, decide candidates
7
7
  export const EDIT_IMG_RE = /^image\d*$/; // edit multi-reference: image, image2, ...
8
8
  export const VID_PORT_RE = /^vid\d+$/;
9
9
  export const CLIP_PORT_RE = /^clip\d+$/; // combine clips
@@ -40,6 +40,7 @@ export const NODE_TYPES = {
40
40
  upload: { title: "Image input", inputs: [], outputs: [{ name: "image", type: "image" }], local: true },
41
41
  aupload: { title: "Audio input", inputs: [], outputs: [{ name: "audio", type: "audio" }], local: true },
42
42
  vupload: { title: "Video input", inputs: [], outputs: [{ name: "video", type: "video" }], local: true },
43
+ mupload: { title: "3D input", inputs: [], outputs: [{ name: "model", type: "model3d" }], local: true },
43
44
  choice: { title: "Choice", inputs: [], outputs: [{ name: "text", type: "text" }], local: true },
44
45
  join: { title: "Join", inputs: ["a", "b"], outputs: [{ name: "text", type: "text" }], local: true },
45
46
  llm: { title: "LLM", inputs: [], outputs: [{ name: "text", type: "text" }], network: true },
@@ -48,8 +49,10 @@ export const NODE_TYPES = {
48
49
  inpaint: { title: "Inpaint", inputs: ["image", "mask"], outputs: [{ name: "image", type: "image" }], network: true },
49
50
  resize: { title: "Resize / crop", inputs: ["image"], outputs: [{ name: "image", type: "image" }], local: true },
50
51
  vision: { title: "Vision", inputs: ["image"], outputs: [{ name: "text", type: "text" }], network: true },
52
+ decide: { title: "Decide", inputs: ["text"], outputs: [{ name: "text", type: "text" }, { name: "image", type: "image" }], network: true }, // img1… candidates (IMG_PORT_RE)
51
53
  tvideo: { title: "Text→Video", inputs: [], outputs: [{ name: "video", type: "video" }], network: true },
52
54
  ivideo: { title: "Image→Video", inputs: ["image"], outputs: [{ name: "video", type: "video" }], network: true },
55
+ model3d: { title: "3D model", inputs: ["image"], outputs: [{ name: "model", type: "model3d" }], network: true },
53
56
  vedit: { title: "Video edit", inputs: ["video"], outputs: [{ name: "video", type: "video" }], network: true },
54
57
  vframes: { title: "Video → frames", inputs: ["video"], outputs: [{ name: "frame1", type: "image" }], local: true, framesOut: true }, // dynamic frame1..N
55
58
  combine: { title: "Combine videos", inputs: [], outputs: [{ name: "video", type: "video" }], local: true },
@@ -61,6 +64,10 @@ export const NODE_TYPES = {
61
64
  trim: { title: "Trim audio", inputs: ["audio"], outputs: [{ name: "audio", type: "audio" }], local: true },
62
65
  extractaudio: { title: "Extract audio", inputs: ["video"], outputs: [{ name: "audio", type: "audio" }], local: true },
63
66
  transcribe: { title: "Transcribe", inputs: ["audio"], outputs: [{ name: "text", type: "text" }], network: true },
67
+ cleanvoice: { title: "Clean voice", inputs: ["audio", "video"], outputs: [{ name: "audio", type: "audio" }], network: true },
68
+ // Custom URL — not a NanoGPT call, so `network` stays unset (no API key required).
69
+ // Output port follows fields.mode (see deriveOutputs / endpointOutPort).
70
+ endpoint: { title: "Custom endpoint", inputs: ["text", "image", "audio", "video"], outputs: [{ name: "text", type: "text" }] },
64
71
  comment: { title: "Comment", inputs: [], outputs: [], note: true, local: true },
65
72
  };
66
73
 
@@ -104,7 +111,10 @@ const MEDIA_FIELD_KEYS = ["image", "mask", "audio", "video"];
104
111
  const MEDIA_URL_RE = /^(data:|https?:)/i;
105
112
 
106
113
  function scrubMediaPlaceholders(n, warnings) {
107
- for (const k of MEDIA_FIELD_KEYS) {
114
+ // mupload stores the .glb on fields.model. That key is an AI model id on every
115
+ // other node, so it is only treated as media here.
116
+ const keys = n.type === "mupload" ? MEDIA_FIELD_KEYS.concat(["model"]) : MEDIA_FIELD_KEYS;
117
+ for (const k of keys) {
108
118
  const v = n.fields[k];
109
119
  if (v == null || v === "" || (typeof v === "string" && MEDIA_URL_RE.test(v.trim()))) continue;
110
120
  const shown = typeof v === "string"
package/src/index.mjs CHANGED
@@ -6,8 +6,12 @@ export {
6
6
  } from "./media.mjs";
7
7
  export { NanoClient } from "./client.mjs";
8
8
  export { NODE_TYPES, displayName, materialize, topoSort, wiredFramesFloor, MAX_FRAMES } from "./graph.mjs";
9
+ // edit models whose image ports are ORDERED roles — a caller building graphs needs the slot order
10
+ export { IMG_INPUT_ROLES } from "./nodes.mjs";
9
11
  export { deriveInputs, deriveOutputs, deriveSettings, INPUT_SPECS, SETTING_SPECS } from "./io.mjs";
10
12
  export { estimateGraphCost, graphModelKinds } from "./estimate.mjs";
13
+ // prompt length caps — a caller doing its own orchestration needs the same trim-and-say-so
14
+ export { PROMPT_CAPS, promptCap, fitPromptText, isPromptTooLong, promptCapFromError } from "./prompt-caps.mjs";
11
15
  export { decodeShareUrl, decodeShareFragment, isShareRef } from "./share.mjs";
12
16
  export { parseNanoInvoice } from "./x402.mjs";
13
17
  export { qrTerminal, qrModules } from "./qr.mjs";
package/src/io.mjs CHANGED
@@ -1,5 +1,6 @@
1
1
  import { NanoodleError } from "./errors.mjs";
2
2
  import { NODE_TYPES, displayName, optionalNode, topoSort, wiredFramesFloor, MAX_FRAMES } from "./graph.mjs";
3
+ import { endpointOutPort } from "./endpoint.mjs";
3
4
 
4
5
  /* ============================== INPUTS ============================== */
5
6
 
@@ -9,6 +10,7 @@ export const INPUT_SPECS = {
9
10
  upload: [{ f: "image", label: "Image", kind: "image" }],
10
11
  aupload: [{ f: "audio", label: "Audio", kind: "audio" }],
11
12
  vupload: [{ f: "video", label: "Video", kind: "video" }],
13
+ mupload: [{ f: "model", label: "3D file", kind: "model3d" }],
12
14
  llm: [{ f: "prompt", label: "Prompt", kind: "textarea" },
13
15
  { f: "system", label: "System prompt", kind: "textarea", optional: true, def: "You are a helpful, concise assistant." }],
14
16
  image: [{ f: "prompt", label: "Image prompt", kind: "textarea" }],
@@ -16,6 +18,9 @@ export const INPUT_SPECS = {
16
18
  music: [{ f: "prompt", label: "Style / prompt", kind: "textarea" }],
17
19
  remix: [{ f: "prompt", label: "Style / direction", kind: "textarea" }],
18
20
  tts: [{ f: "prompt", label: "Text to speak", kind: "textarea" }],
21
+ // Optional: image-only 3D models run with a photo and no prompt.
22
+ model3d: [{ f: "prompt", label: "3D prompt", kind: "textarea", optional: true }],
23
+ endpoint:[{ f: "prompt", label: "Prompt", kind: "textarea" }],
19
24
  };
20
25
 
21
26
  /**
@@ -163,6 +168,8 @@ export function deriveOutputs(graph) {
163
168
  ports = [];
164
169
  for (let i = 1; i <= count; i++) ports.push({ name: "frame" + i, type: "image" });
165
170
  }
171
+ // Custom endpoint's output port follows fields.mode (chat→text, image, video, audio, json→text).
172
+ if (n.type === "endpoint") ports = [endpointOutPort(n)];
166
173
  return { key, nodeId: n.id, type: n.type, ports };
167
174
  });
168
175
  }
@@ -187,6 +194,14 @@ export const SETTING_SPECS = {
187
194
  { f: "model", label: "Model", kind: "model" },
188
195
  { f: "q", label: "Question", kind: "textarea", def: "Describe this image." },
189
196
  ],
197
+ decide: [
198
+ { f: "model", label: "Model", kind: "model" },
199
+ { f: "mode", label: "Decide", kind: "select", options: ["pick", "choose", "score", "yesno"], def: "pick" },
200
+ { f: "question", label: "Question", kind: "textarea" },
201
+ { f: "options", label: "Labels (one per line)", kind: "textarea" },
202
+ { f: "levels", label: "Scale (worst → best, one per line)", kind: "textarea", def: "poor\nokay\ngood\ngreat" },
203
+ { f: "gate", label: "Gate: stop downstream on no", kind: "boolean" },
204
+ ],
190
205
  image: [
191
206
  { f: "model", label: "Model", kind: "model" },
192
207
  { f: "size", label: "Image size", kind: "select", options: SIZES, def: "1024x1024" },
@@ -245,6 +260,19 @@ export const SETTING_SPECS = {
245
260
  { f: "model", label: "Model", kind: "model" },
246
261
  { f: "language", label: "Language", kind: "text", def: "auto" },
247
262
  ],
263
+ cleanvoice: [
264
+ { f: "model", label: "Model", kind: "model" },
265
+ { f: "url", label: "Public link (when nothing is wired)", kind: "text" },
266
+ ],
267
+ model3d: [
268
+ { f: "model", label: "Model", kind: "model" },
269
+ ],
270
+ endpoint: [
271
+ { f: "url", label: "URL", kind: "text", def: "http://127.0.0.1:8787/v1/chat/completions" },
272
+ { f: "mode", label: "Mode", kind: "select", options: ["chat", "image", "video", "audio", "json"], def: "chat" },
273
+ { f: "model", label: "Model", kind: "text", def: "local" },
274
+ { f: "auth", label: "Authorization (optional)", kind: "text", def: "" },
275
+ ],
248
276
  join: [{ f: "sep", label: "Separator (use \\n for a line break)", kind: "text", def: " " }],
249
277
  inpaint: [
250
278
  { f: "model", label: "Model", kind: "model" },
@@ -729,6 +729,44 @@ export async function fitImageInline(url, { budget = INLINE_IMAGE_BUDGET, fetch:
729
729
  throw new NanoodleError("image is too large to send inline even after resizing (~4 MB max) — use a smaller image");
730
730
  }
731
731
 
732
+ /**
733
+ * Shrink an image to fit a decision model's limits (⚖️ Decide): long edge ≤ maxDim and the
734
+ * data: URL ≤ budget characters. JPEG on white (a transparent PNG can't turn black), stepping
735
+ * quality down, then size, until it fits — the twin of the browser's canvas decideFitImage.
736
+ * Needs ffmpeg (decisions without images never touch this).
737
+ */
738
+ export async function fitImageJpeg(url, maxDim, budget, { fetch: fetchFn, signal } = {}) {
739
+ return withTemp(async (dir) => {
740
+ throwIfAborted(signal);
741
+ const inPath = await writeInput(dir, "in", url, fetchFn);
742
+ const probe = await runProc("ffprobe", [
743
+ "-v", "error", "-select_streams", "v:0",
744
+ "-show_entries", "stream=width,height", "-of", "csv=p=0:s=x", inPath,
745
+ ], { signal });
746
+ const dims = String(probe.stdout).trim().split("x").map(Number);
747
+ const sw = dims[0], sh = dims[1];
748
+ if (!(sw > 0) || !(sh > 0)) throw new NanoodleError("couldn’t read an image to judge");
749
+ let s = Math.min(1, maxDim / Math.max(sw, sh));
750
+ const qs = [3, 5, 8, 12]; // mjpeg -q:v (2 = best … 31 = worst) ≈ canvas quality 0.85 → 0.48
751
+ for (let round = 0; round < 6; round++) {
752
+ const w = Math.max(2, Math.round(sw * s / 2) * 2), h = Math.max(2, Math.round(sh * s / 2) * 2);
753
+ for (const q of qs) {
754
+ throwIfAborted(signal);
755
+ const outPath = join(dir, `fit-${round}-${q}.jpg`);
756
+ await runProc("ffmpeg", [
757
+ "-y", "-i", inPath, "-filter_complex",
758
+ `color=c=white:s=${w}x${h}[bg];[0:v]scale=${w}:${h},format=rgba[fg];[bg][fg]overlay=shortest=1,format=yuvj420p`,
759
+ "-frames:v", "1", "-q:v", String(q), outPath,
760
+ ], { signal });
761
+ const out = await dataUrlFromFile(outPath, "image/jpeg");
762
+ if (out.length <= budget) return out;
763
+ }
764
+ s *= 0.75;
765
+ }
766
+ throw new NanoodleError("couldn’t shrink an image small enough for the decision model");
767
+ });
768
+ }
769
+
732
770
  async function resizeCropImageFfmpeg(url, m, w, h, { fetch: fetchFn, signal } = {}) {
733
771
  return withTemp(async (dir) => {
734
772
  throwIfAborted(signal);
package/src/media.mjs CHANGED
@@ -42,6 +42,7 @@ const EXT_MIME = {
42
42
  ".mp3": "audio/mpeg", ".wav": "audio/wav", ".ogg": "audio/ogg", ".opus": "audio/ogg",
43
43
  ".aac": "audio/aac", ".flac": "audio/flac", ".m4a": "audio/mp4",
44
44
  ".mp4": "video/mp4", ".webm": "video/webm", ".mov": "video/quicktime", ".mkv": "video/x-matroska",
45
+ ".glb": "model/gltf-binary",
45
46
  ".txt": "text/plain", ".json": "application/json",
46
47
  };
47
48
 
@@ -50,6 +51,7 @@ const MIME_EXT = {
50
51
  "audio/mpeg": "mp3", "audio/wav": "wav", "audio/ogg": "ogg", "audio/aac": "aac",
51
52
  "audio/flac": "flac", "audio/mp4": "m4a",
52
53
  "video/mp4": "mp4", "video/webm": "webm", "video/quicktime": "mov",
54
+ "model/gltf-binary": "glb",
53
55
  "text/plain": "txt", "application/json": "json",
54
56
  };
55
57
 
@@ -80,6 +82,7 @@ export function sniffMime(bytes) {
80
82
  if (ascii(0, "fLaC")) return "audio/flac";
81
83
  if (b.length >= 12 && ascii(4, "ftyp")) return "video/mp4";
82
84
  if (b.length >= 4 && b[0] === 0x1a && b[1] === 0x45 && b[2] === 0xdf && b[3] === 0xa3) return "video/webm";
85
+ if (b.length >= 4 && ascii(0, "glTF")) return "model/gltf-binary";
83
86
  return "application/octet-stream";
84
87
  }
85
88
 
package/src/nodes.mjs CHANGED
@@ -1,11 +1,14 @@
1
1
  import { NanoodleError } from "./errors.mjs";
2
- import { catItem, chatModelCan } from "./catalog.mjs";
2
+ import { catItem, chatModelCan, pricingAdvertisesRefs } from "./catalog.mjs";
3
3
  import { IMG_PORT_RE, EDIT_IMG_RE, REF_PORT_RE, CLIP_PORT_RE, VID_PORT_RE, optionalNode } from "./graph.mjs";
4
4
  import { MEDIA_INLINE_MAX } from "./media.mjs";
5
5
  import {
6
6
  resizeCropImage, trimAudioToWav, extractAudioToWav,
7
- extractVideoFrames, concatVideos, muxSoundtrack, maskToSource, fitImageInline,
7
+ extractVideoFrames, concatVideos, muxSoundtrack, maskToSource, fitImageInline, fitImageJpeg,
8
8
  } from "./local-media.mjs";
9
+ import { decideRun, decideImageLimits } from "./decide.mjs";
10
+ import { cleanVoiceRun, CLEANVOICE_DEFAULT_MODEL } from "./cleanvoice.mjs";
11
+ import { runEndpoint } from "./endpoint.mjs";
9
12
 
10
13
  function mdl(n) {
11
14
  const m = String((n.fields && n.fields.model) || "").trim();
@@ -52,6 +55,46 @@ function collectPorts(inp, re) {
52
55
  .filter(Boolean);
53
56
  }
54
57
 
58
+ /**
59
+ * Edit models that demand a FIXED, ORDERED set of input images. Nothing in the live catalog
60
+ * declares a minimum — a full 215-model scan (2026-07-31) found no min_items/min_input_images
61
+ * field, only free text — so the roles live here. The order IS the slot order: role 1 = the
62
+ * `image` port, role 2 = `image2`, … (flux vto auto-detects the wearer either way, but a
63
+ * garment-first send drifts the identity toward the garment photo's model).
64
+ */
65
+ export const IMG_INPUT_ROLES = {
66
+ "flux-pro/v1/vto": ["person", "garment"],
67
+ };
68
+
69
+ const imgSlot = (i) => (i === 1 ? "image" : "image" + i); // slot index → EDIT_IMG_RE port name
70
+
71
+ /**
72
+ * Refuse a role-model edit BEFORE the paid call unless every slot is wired. Read the RAW inp
73
+ * keys, never collectPorts: that compacts (.filter(Boolean)), so wiring only `image2` promotes
74
+ * the garment into the person slot and buys a silently wrong result.
75
+ */
76
+ function guardInputRoles(model, inp) {
77
+ const roles = IMG_INPUT_ROLES[String(model || "").trim()];
78
+ if (!roles) return;
79
+ const missing = roles.filter((r, i) => !inp[imgSlot(i + 1)]);
80
+ if (!missing.length) return;
81
+ const slots = roles.map((r, i) => `${r} (${imgSlot(i + 1)})`).join(", ");
82
+ throw new NanoodleError(
83
+ `${model} needs ${roles.length} images: ${slots} — ${roles.length - missing.length} wired (missing: ${missing.join(", ")})`,
84
+ );
85
+ }
86
+
87
+ /**
88
+ * edit/inpaint have a single image output, but some models always bill (and return) a fixed
89
+ * batch — midjourney/text-to-image and higgsfield-soul declare fixed_image_count: 4. Ask with
90
+ * multi so we can SEE the extras and say we dropped them; the return shape stays one url.
91
+ */
92
+ function keptFirst(urls, ctx) {
93
+ if (!Array.isArray(urls)) return urls; // host-injected ctx.image that ignores multi → already one url
94
+ if (urls.length > 1 && ctx && ctx.progress) ctx.progress(`model returned ${urls.length} images, kept the first`);
95
+ return urls[0];
96
+ }
97
+
55
98
  function promptOf(n, inp, errMsg) {
56
99
  const raw = inp.prompt != null ? inp.prompt : n.fields.prompt != null ? n.fields.prompt : "";
57
100
  const p = String(raw).trim();
@@ -126,7 +169,10 @@ function loraFamily(model) {
126
169
  if (/flux-2/i.test(m)) return "flux2dev";
127
170
  if (/z-image/i.test(m)) return "zimage";
128
171
  if (/ltx/i.test(m)) return "ltx";
129
- if (/lora/i.test(m)) return "flux";
172
+ if (/minimax-h3/i.test(m)) return "h3";
173
+ if (/anima/i.test(m) && /lora/i.test(m)) return "anima";
174
+ if (/qwen-image-2\.1/i.test(m) && /lora/i.test(m)) return "qwen21";
175
+ if (/lora/i.test(m)) return /krea/i.test(m) ? "krea" : "flux";
130
176
  return null;
131
177
  }
132
178
 
@@ -136,6 +182,7 @@ function imageTakesLora(id) {
136
182
  id = String(id || "");
137
183
  if (/inpaint/i.test(id)) return false;
138
184
  if (/klein/i.test(id)) return true;
185
+ if (/minimax-h3\/(text-to-image|image-edit)/i.test(id)) return true;
139
186
  return /(^|[-\/])lora($|[-\/])/i.test(id);
140
187
  }
141
188
 
@@ -149,7 +196,7 @@ function modelTakesLora(kind, id) {
149
196
  function loraCap(model) {
150
197
  switch (loraFamily(model)) {
151
198
  case "flux2dev": return 4;
152
- case "flux2klein": case "zimage": case "ltx": return 3;
199
+ case "flux2klein": case "zimage": case "ltx": case "krea": case "h3": case "anima": case "qwen21": return 3;
153
200
  default: return 1; // flux-lora, pimage — single slot
154
201
  }
155
202
  }
@@ -165,7 +212,7 @@ function nodeLoras(n) {
165
212
  function loraBodyFor(model, items) {
166
213
  const fam = loraFamily(model), sc = (v) => (isNaN(v) ? 1 : v);
167
214
  if (fam === "pimage") return { lora_weights: items[0].url, lora_scale: sc(items[0].scale) };
168
- if (fam === "flux2dev" || fam === "flux2klein" || fam === "zimage" || fam === "ltx") {
215
+ if (fam === "flux2dev" || fam === "flux2klein" || fam === "zimage" || fam === "ltx" || fam === "krea" || fam === "h3" || fam === "anima" || fam === "qwen21") {
169
216
  const b = {};
170
217
  items.forEach((it, i) => { b["lora_url_" + (i + 1)] = it.url; b["lora_scale_" + (i + 1)] = sc(it.scale); });
171
218
  return b;
@@ -237,6 +284,13 @@ function imgExtra(n) {
237
284
  const np = String(n.fields.negativePrompt || "").trim();
238
285
  if (np) e.negative_prompt = np;
239
286
  }
287
+ // FIBO 1.5: size is 1mp/4mp; shape rides as aspect_ratio. Only this advertised id —
288
+ // leftover fields.aspect after a model swap must not go out on Muse/Recraft/etc.
289
+ // Marketing /api/models additionalParams.aspect_ratio (v1 image-models still omits it).
290
+ if (n.fields.model === "bria/fibo-generate-1.5/text-to-image") {
291
+ const av = (n.fields.aspect != null && n.fields.aspect !== "") ? n.fields.aspect : "1:1";
292
+ if (av !== "") e.aspect_ratio = String(av);
293
+ }
240
294
  return e;
241
295
  }
242
296
 
@@ -255,12 +309,17 @@ function videoDims(n, ctx) {
255
309
  const aWire = p.aspect_ratio ? "aspect_ratio" : p.orientation ? "orientation" : p.resolution_ratio ? "resolution_ratio" : "aspect_ratio";
256
310
  const dP = p.duration || p.seconds;
257
311
  const dWire = p.duration ? "duration" : p.seconds ? "seconds" : "duration";
312
+ // tvideo/ivideo soft-send aspect+duration (those nodes always show the knobs).
313
+ // vedit/lipsync and resolution never forward a leftover the model did not list.
314
+ // Gallery: Omni v1 has duration+aspect only; leftover 720p/4k from v1.1 must not POST.
315
+ // Matches play videoDimParams (pack.resolution / dimDefs) — including catalog-miss omit.
316
+ const soft = n.type !== "vedit" && n.type !== "lipsync";
258
317
  let asp = f.aspect, dur = f.duration;
259
318
  if ((asp == null || asp === "") && aP && aP.default != null) asp = aP.default;
260
319
  if ((dur == null || dur === "") && dP && dP.default != null) dur = dP.default;
261
- if (f.resolution != null && f.resolution !== "") out.resolution = f.resolution;
262
- if (asp != null && asp !== "") out[aWire] = asp;
263
- if (dur != null && dur !== "") out[dWire] = dur;
320
+ if (p.resolution && f.resolution != null && f.resolution !== "") out.resolution = f.resolution;
321
+ if ((aP || soft) && asp != null && asp !== "") out[aWire] = asp;
322
+ if ((dP || soft) && dur != null && dur !== "") out[dWire] = dur;
264
323
  return out;
265
324
  }
266
325
 
@@ -268,10 +327,13 @@ function videoDims(n, ctx) {
268
327
  Video models disagree on the ref-array param name AND its size limit; sending the wrong
269
328
  key silently degrades to a plain video, sending too many can over-bill. Resolve the
270
329
  model's REAL key from the catalog and clamp to its declared max. */
271
- function refMaxFor(model) {
330
+ function refMaxFor(model, pricing) {
272
331
  const id = String(model || "");
273
332
  if (/seedance/i.test(id)) return 9;
333
+ if (/minimax-h3/i.test(id)) return 9; // upstream cap; the first 5 are included, 6-9 bill extra_reference_image each
274
334
  if (/luma|ray/i.test(id)) return 4;
335
+ const inc = pricing && +pricing.included_reference_images;
336
+ if (inc > 0) return inc; // no family entry: the included count is the only number the catalog gives us
275
337
  return 4;
276
338
  }
277
339
 
@@ -285,15 +347,15 @@ function modelRefSpec(model, ctx) {
285
347
  const m = catItem(ctx && ctx.catalog, "video", model);
286
348
  if (!m) return { key: "reference_images", cap: refMaxFor(model) };
287
349
  const sp = m.supported_parameters || {}, pp = sp.parameters || sp;
288
- const key = keys.find((k) => k in pp);
289
- if (!key) return null; // known model with no ref-image param
350
+ const key = keys.find((k) => k in pp) || (pricingAdvertisesRefs(m.pricing) ? "reference_images" : null);
351
+ if (!key) return null; // known model with no ref-image param and no ref pricing
290
352
  const d = pp[key];
291
353
  let cap = null;
292
354
  if (d && typeof d === "object") {
293
355
  const mx = d.max != null ? d.max : d.maxItems != null ? d.maxItems : d.max_items;
294
356
  if (mx != null && +mx > 0) cap = +mx;
295
357
  }
296
- return { key, cap: cap != null ? cap : refMaxFor(model) };
358
+ return { key, cap: cap != null ? cap : refMaxFor(model, m.pricing) };
297
359
  }
298
360
 
299
361
  /** Attach wired refs to video opts under the model's real key, clamped to its cap (twin of the app runtimes: say so, never silently discard). */
@@ -337,22 +399,42 @@ function audioParams(n, ctx) {
337
399
  const durOk = !sp || (sp.min_duration != null && sp.max_duration != null); // cat:duration
338
400
  const secDurOk = !sp || (sp.min_duration != null && sp.max_duration != null && +((m.pricing || {}).per_second) > 0); // cat:secduration
339
401
  const voiceOk = !sp || (Array.isArray(sp.voices) && sp.voices.length > 0); // cat:voices
402
+ // Yue2 catalogs the file type as output_format (mp3/wav/flac). The OpenAI response_format
403
+ // key is ignored, so a wav/flac choice came back as mp3. No output_format in the catalog
404
+ // (or no catalog) keeps the old response_format send.
405
+ const putFormat = (fmt) => {
406
+ if (!nonEmpty(fmt)) return;
407
+ const of = sp && sp.output_format;
408
+ const values = of && Array.isArray(of.values) ? of.values : null;
409
+ if (values) {
410
+ if (values.indexOf(fmt) < 0) return;
411
+ const def = of.default != null ? of.default : "mp3";
412
+ if (fmt === def) return;
413
+ body.output_format = fmt;
414
+ return;
415
+ }
416
+ if (fmt !== "mp3") body.response_format = fmt;
417
+ };
340
418
  if (n.type === "music") {
341
419
  if (nonEmpty(f.lyrics)) body.lyrics = f.lyrics;
342
420
  if (f.instrumental === true || f.instrumental === "true") body.instrumental = true;
343
421
  if (nonEmpty(f.duration) && num(f.duration) != null && durOk) body.duration = num(f.duration);
344
422
  if (nonEmpty(f.negative_prompt)) body.negative_prompt = f.negative_prompt;
345
423
  if (nonEmpty(f.seed) && num(f.seed) != null) body.seed = num(f.seed);
346
- if (nonEmpty(f.response_format) && f.response_format !== "mp3") body.response_format = f.response_format;
424
+ putFormat(f.response_format);
347
425
  } else if (n.type === "tts") {
348
426
  if (nonEmpty(f.voice) && voiceOk) body.voice = f.voice;
349
427
  if (nonEmpty(f.speed) && num(f.speed) != null && num(f.speed) !== 1) body.speed = num(f.speed); // omit when 1
350
428
  if (nonEmpty(f.instructions)) body.instructions = f.instructions;
351
- if (nonEmpty(f.response_format) && f.response_format !== "mp3") body.response_format = f.response_format;
429
+ putFormat(f.response_format);
352
430
  } else if (n.type === "remix") {
353
431
  if (nonEmpty(f.lyrics)) body.lyrics = f.lyrics;
354
432
  if (nonEmpty(f.duration) && num(f.duration) != null && secDurOk) body.duration = num(f.duration);
355
- if (nonEmpty(f.response_format) && f.response_format !== "mp3") body.response_format = f.response_format;
433
+ // Format is a remix knob only when the catalog lists output_format (Yue2 music-to-music).
434
+ // A cover model with no such key must not receive a stale response_format.
435
+ const of = sp && sp.output_format;
436
+ const fmtListed = !!(of && Array.isArray(of.values) && of.values.length);
437
+ if (!sp || fmtListed) putFormat(f.response_format);
356
438
  }
357
439
  if ((f.extraJson || "").trim()) {
358
440
  try { Object.assign(body, JSON.parse(f.extraJson)); }
@@ -422,6 +504,13 @@ export const RUNNERS = {
422
504
  }
423
505
  return { video: n.fields.video };
424
506
  },
507
+ async mupload(n) {
508
+ if (!n.fields.model) {
509
+ if (optionalNode(n)) return { model: "" };
510
+ throw new NanoodleError("no 3D file — drop a .glb first");
511
+ }
512
+ return { model: n.fields.model };
513
+ },
425
514
 
426
515
  async choice(n) {
427
516
  const opts = String(n.fields.options || "").split("\n").map((s) => s.trim()).filter(Boolean);
@@ -504,7 +593,16 @@ export const RUNNERS = {
504
593
 
505
594
  async llm(n, inp, ctx) {
506
595
  const prompt = promptOf(n, inp, "no prompt");
507
- const imgs = await Promise.all(collectPorts(inp, IMG_PORT_RE).map((u) => fitImage(u, ctx, "wired image")));
596
+ // Wired image ports: a text-only chat model can't see the pictures and NanoGPT still
597
+ // bills the image tokens, so drop them for a KNOWN text-only model and note it —
598
+ // mirrors play's chatModelCan("vision") / editor modelSupportsImages gate (permissive
599
+ // for catalog-absent models; twin of the audio gate below).
600
+ const allImgs = collectPorts(inp, IMG_PORT_RE);
601
+ const seesImages = chatModelCan(ctx.catalog, mdl(n), "vision");
602
+ const imgs = seesImages
603
+ ? await Promise.all(allImgs.map((u) => fitImage(u, ctx, "wired image")))
604
+ : [];
605
+ if (allImgs.length && !seesImages) ctx.progress("image(s) ignored — this model is text-only");
508
606
  // hosted audio (music/tts nodes return https CDN URLs verbatim) → download + inline as base64:
509
607
  // the chat input_audio part carries bytes, never a URL
510
608
  let audioPart = null;
@@ -536,6 +634,17 @@ export const RUNNERS = {
536
634
  return { text: await ctx.chat(messages, mdl(n), {}) };
537
635
  },
538
636
 
637
+ // ⚖️ Decide: one typed question to a NanoGPT decision model (twin of the editor/play node).
638
+ // A closed yes/no gate throws a NanoodleError with code "decide-gate" — downstream skips, unbilled.
639
+ async decide(n, inp, ctx) {
640
+ const model = mdl(n);
641
+ const it = catItem(ctx.catalog, "chat", model);
642
+ // a catalog row WITHOUT decision_input (e.g. the non-detailed /api/v1/models list) is "unknown", not text-only
643
+ const lim = decideImageLimits(it && it.decision_input, !!(it && it.decision_input));
644
+ const fit = (u, maxDim, budget) => fitImageJpeg(u, maxDim, budget, mediaOpts(ctx));
645
+ return decideRun(n.fields || {}, model, inp.text, collectPorts(inp, IMG_PORT_RE), lim, (body) => ctx.decide(body), fit);
646
+ },
647
+
539
648
  async image(n, inp, ctx) {
540
649
  const prompt = promptOf(n, inp, "no prompt");
541
650
  let want = Math.max(1, parseInt(n.fields.variations, 10) || 1);
@@ -549,6 +658,7 @@ export const RUNNERS = {
549
658
  },
550
659
 
551
660
  async edit(n, inp, ctx) {
661
+ guardInputRoles(n.fields.model, inp); // role models: refuse holes before spending
552
662
  let imgs = collectPorts(inp, EDIT_IMG_RE);
553
663
  if (!imgs.length) throw new NanoodleError("no image input");
554
664
  // cap to the model's max_input_images (item present but silent → 1; absent → no cap):
@@ -567,7 +677,8 @@ export const RUNNERS = {
567
677
  imgs = await Promise.all(imgs.map((u) => fitImage(u, ctx, "source image")));
568
678
  guardRefsSize(imgs);
569
679
  const src = imgs.length > 1 ? imgs : imgs[0]; // array → multi-image composite; string → single edit
570
- return { image: await ctx.image({ prompt, model: mdl(n), size: n.fields.size || "1024x1024", imageDataUrl: src, extra: imgExtra(n) }) };
680
+ const urls = await ctx.image({ prompt, model: mdl(n), size: n.fields.size || "1024x1024", imageDataUrl: src, extra: imgExtra(n), multi: true });
681
+ return { image: keptFirst(urls, ctx) };
571
682
  },
572
683
 
573
684
  async inpaint(n, inp, ctx) {
@@ -580,7 +691,8 @@ export const RUNNERS = {
580
691
  // ctx.maskToSource lets a browser host inject its canvas compositor (handles JPEG/WebP
581
692
  // sources the pure-PNG path can't, where ffmpeg isn't an option).
582
693
  const mask = await (ctx.maskToSource || maskToSource)(rawMask, source, mediaOpts(ctx));
583
- return { image: await ctx.image({ prompt, model: mdl(n), size: n.fields.size || "1024x1024", imageDataUrl: source, maskDataUrl: mask, extra: imgExtra(n) }) };
694
+ const urls = await ctx.image({ prompt, model: mdl(n), size: n.fields.size || "1024x1024", imageDataUrl: source, maskDataUrl: mask, extra: imgExtra(n), multi: true });
695
+ return { image: keptFirst(urls, ctx) };
584
696
  },
585
697
 
586
698
  async tvideo(n, inp, ctx) {
@@ -688,4 +800,58 @@ export const RUNNERS = {
688
800
  if (!inp.audio) throw new NanoodleError("no audio input");
689
801
  return { text: await ctx.transcribe(mdl(n), inp.audio, (n.fields.language || "auto").trim()) };
690
802
  },
803
+
804
+ // 🎧 Clean voice: hosted audio or video → voice-only audio. data:/blob: refused before any request.
805
+ async cleanvoice(n, inp, ctx) {
806
+ const model = String((n.fields && n.fields.model) || "").trim() || CLEANVOICE_DEFAULT_MODEL;
807
+ return cleanVoiceRun(model, inp, n.fields, (m, extra) => ctx.audio(m, "", extra));
808
+ },
809
+
810
+ // 🧊 3D model: image and/or text → GLB via the video submit/poll API (mediaKind model3d).
811
+ async model3d(n, inp, ctx) {
812
+ const model = String((n.fields && n.fields.model) || "").trim() || MODEL3D_IMAGE_DEFAULT;
813
+ const mods = model3dInputMods(ctx && ctx.catalog, model);
814
+ const prompt = promptOf(n, inp);
815
+ let image = mods.image ? (inp.image || "") : "";
816
+ if (inp.image && !mods.image && ctx && ctx.progress) {
817
+ ctx.progress("This model takes text only — the image wire stays, but this run ignores it.");
818
+ }
819
+ if (prompt && !mods.text && ctx && ctx.progress) {
820
+ ctx.progress("This model takes an image only — the prompt stays, but this run ignores it.");
821
+ }
822
+ if (mods.image && !mods.text && !image) throw new NanoodleError("Connect an image first.");
823
+ if (mods.text && !mods.image && !prompt) throw new NanoodleError("Add a prompt first.");
824
+ if (mods.image && mods.text && !image && !prompt) throw new NanoodleError("Add a prompt or connect an image first.");
825
+ if (image) image = await fitImage(image, ctx, "source image");
826
+ const url = await ctx.video(model, mods.text ? prompt : "", {
827
+ extra: n.fields.modelOpts || {},
828
+ mediaKind: "model3d",
829
+ }, image || null);
830
+ return { model: url };
831
+ },
832
+
833
+ // 🔌 Custom endpoint: POST to the graph's own URL. ctx.fetch so tests can point it at a mock.
834
+ async endpoint(n, inp, ctx) {
835
+ return runEndpoint(n, inp, { fetch: ctx && ctx.fetch, signal: ctx && ctx.signal });
836
+ },
691
837
  };
838
+
839
+ // Image→3D (Tripo) takes a photo only. Hunyuan Rapid takes text and an optional image.
840
+ // Anything else, including a catalog miss, accepts either — the editor's same fallback.
841
+ const MODEL3D_IMAGE_DEFAULT = "tripo3d/v2.5";
842
+ const MODEL3D_TEXT_DEFAULT = "wavespeed-ai/hunyuan-3d-v3.1-rapid";
843
+
844
+ function model3dInputMods(catalog, id) {
845
+ const it = catItem(catalog, "model3d", id);
846
+ const raw = it && (
847
+ (Array.isArray(it.modalities) && it.modalities)
848
+ || (it.architecture && Array.isArray(it.architecture.input_modalities) && it.architecture.input_modalities)
849
+ );
850
+ if (raw && raw.length) {
851
+ const mods = raw.filter((x) => x === "image" || x === "text");
852
+ if (mods.length) return { image: mods.includes("image"), text: mods.includes("text") };
853
+ }
854
+ if (id === MODEL3D_IMAGE_DEFAULT) return { image: true, text: false };
855
+ if (id === MODEL3D_TEXT_DEFAULT) return { image: true, text: true };
856
+ return { image: true, text: true };
857
+ }