nanoodle 0.8.0 → 0.11.0

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
 
@@ -83,6 +90,12 @@ export function optionalNode(node) {
83
90
  return v === true || v === "true";
84
91
  }
85
92
 
93
+ /** Author-locked node: fields.locked keeps every field at its stored value (no inputs derived). */
94
+ export function lockedNode(node) {
95
+ const v = node && node.fields && node.fields.locked;
96
+ return v === true || v === "true";
97
+ }
98
+
86
99
  /** Is a wire landing on `port` of `node` a data input (vs a field override)? */
87
100
  export function isInputPort(node, port) {
88
101
  const t = NODE_TYPES[node.type];
@@ -104,7 +117,10 @@ const MEDIA_FIELD_KEYS = ["image", "mask", "audio", "video"];
104
117
  const MEDIA_URL_RE = /^(data:|https?:)/i;
105
118
 
106
119
  function scrubMediaPlaceholders(n, warnings) {
107
- for (const k of MEDIA_FIELD_KEYS) {
120
+ // mupload stores the .glb on fields.model. That key is an AI model id on every
121
+ // other node, so it is only treated as media here.
122
+ const keys = n.type === "mupload" ? MEDIA_FIELD_KEYS.concat(["model"]) : MEDIA_FIELD_KEYS;
123
+ for (const k of keys) {
108
124
  const v = n.fields[k];
109
125
  if (v == null || v === "" || (typeof v === "string" && MEDIA_URL_RE.test(v.trim()))) continue;
110
126
  const shown = typeof v === "string"
package/src/index.mjs CHANGED
@@ -6,6 +6,8 @@ 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";
11
13
  // prompt length caps — a caller doing its own orchestration needs the same trim-and-say-so
package/src/io.mjs CHANGED
@@ -1,5 +1,6 @@
1
1
  import { NanoodleError } from "./errors.mjs";
2
- import { NODE_TYPES, displayName, optionalNode, topoSort, wiredFramesFloor, MAX_FRAMES } from "./graph.mjs";
2
+ import { NODE_TYPES, displayName, lockedNode, 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
  /**
@@ -36,6 +41,9 @@ export function deriveInputs(graph) {
36
41
  };
37
42
  for (const n of graph.nodes) {
38
43
  if (n.unknown) continue;
44
+ // fields.locked: the node's fields are fixed by the graph — never surfaced as inputs,
45
+ // so runs use the stored values.
46
+ if (lockedNode(n)) continue;
39
47
  if (n.type === "inpaint") {
40
48
  if (!fed(n.id, "prompt")) entries.push(mk(n, "prompt", "What to paint in", "textarea", false));
41
49
  // image and/or mask surface whenever not wired (SPEC-io). The app's combined brush widget
@@ -163,6 +171,8 @@ export function deriveOutputs(graph) {
163
171
  ports = [];
164
172
  for (let i = 1; i <= count; i++) ports.push({ name: "frame" + i, type: "image" });
165
173
  }
174
+ // Custom endpoint's output port follows fields.mode (chat→text, image, video, audio, json→text).
175
+ if (n.type === "endpoint") ports = [endpointOutPort(n)];
166
176
  return { key, nodeId: n.id, type: n.type, ports };
167
177
  });
168
178
  }
@@ -187,6 +197,14 @@ export const SETTING_SPECS = {
187
197
  { f: "model", label: "Model", kind: "model" },
188
198
  { f: "q", label: "Question", kind: "textarea", def: "Describe this image." },
189
199
  ],
200
+ decide: [
201
+ { f: "model", label: "Model", kind: "model" },
202
+ { f: "mode", label: "Decide", kind: "select", options: ["pick", "choose", "score", "yesno"], def: "pick" },
203
+ { f: "question", label: "Question", kind: "textarea" },
204
+ { f: "options", label: "Labels (one per line)", kind: "textarea" },
205
+ { f: "levels", label: "Scale (worst → best, one per line)", kind: "textarea", def: "poor\nokay\ngood\ngreat" },
206
+ { f: "gate", label: "Gate: stop downstream on no", kind: "boolean" },
207
+ ],
190
208
  image: [
191
209
  { f: "model", label: "Model", kind: "model" },
192
210
  { f: "size", label: "Image size", kind: "select", options: SIZES, def: "1024x1024" },
@@ -245,6 +263,19 @@ export const SETTING_SPECS = {
245
263
  { f: "model", label: "Model", kind: "model" },
246
264
  { f: "language", label: "Language", kind: "text", def: "auto" },
247
265
  ],
266
+ cleanvoice: [
267
+ { f: "model", label: "Model", kind: "model" },
268
+ { f: "url", label: "Public link (when nothing is wired)", kind: "text" },
269
+ ],
270
+ model3d: [
271
+ { f: "model", label: "Model", kind: "model" },
272
+ ],
273
+ endpoint: [
274
+ { f: "url", label: "URL", kind: "text", def: "http://127.0.0.1:8787/v1/chat/completions" },
275
+ { f: "mode", label: "Mode", kind: "select", options: ["chat", "image", "video", "audio", "json"], def: "chat" },
276
+ { f: "model", label: "Model", kind: "text", def: "local" },
277
+ { f: "auth", label: "Authorization (optional)", kind: "text", def: "" },
278
+ ],
248
279
  join: [{ f: "sep", label: "Separator (use \\n for a line break)", kind: "text", def: " " }],
249
280
  inpaint: [
250
281
  { 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
 
@@ -135,7 +181,10 @@ function loraKind(type) { return (type === "image" || type === "edit" || type ==
135
181
  function imageTakesLora(id) {
136
182
  id = String(id || "");
137
183
  if (/inpaint/i.test(id)) return false;
138
- if (/klein/i.test(id)) return true;
184
+ // flux-2-klein-* only (lora-capable per catalog); FLUX.3 Klein (blackforestlabs/flux-3/klein/…)
185
+ // takes no lora_url — twin of the app's imageTakesLora (nanoodle #732)
186
+ if (/flux-2-klein/i.test(id)) return true;
187
+ if (/minimax-h3\/(text-to-image|image-edit)/i.test(id)) return true;
139
188
  return /(^|[-\/])lora($|[-\/])/i.test(id);
140
189
  }
141
190
 
@@ -149,7 +198,7 @@ function modelTakesLora(kind, id) {
149
198
  function loraCap(model) {
150
199
  switch (loraFamily(model)) {
151
200
  case "flux2dev": return 4;
152
- case "flux2klein": case "zimage": case "ltx": return 3;
201
+ case "flux2klein": case "zimage": case "ltx": case "krea": case "h3": case "anima": case "qwen21": return 3;
153
202
  default: return 1; // flux-lora, pimage — single slot
154
203
  }
155
204
  }
@@ -165,7 +214,7 @@ function nodeLoras(n) {
165
214
  function loraBodyFor(model, items) {
166
215
  const fam = loraFamily(model), sc = (v) => (isNaN(v) ? 1 : v);
167
216
  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") {
217
+ if (fam === "flux2dev" || fam === "flux2klein" || fam === "zimage" || fam === "ltx" || fam === "krea" || fam === "h3" || fam === "anima" || fam === "qwen21") {
169
218
  const b = {};
170
219
  items.forEach((it, i) => { b["lora_url_" + (i + 1)] = it.url; b["lora_scale_" + (i + 1)] = sc(it.scale); });
171
220
  return b;
@@ -237,6 +286,13 @@ function imgExtra(n) {
237
286
  const np = String(n.fields.negativePrompt || "").trim();
238
287
  if (np) e.negative_prompt = np;
239
288
  }
289
+ // FIBO 1.5: size is 1mp/4mp; shape rides as aspect_ratio. Only this advertised id —
290
+ // leftover fields.aspect after a model swap must not go out on Muse/Recraft/etc.
291
+ // Marketing /api/models additionalParams.aspect_ratio (v1 image-models still omits it).
292
+ if (n.fields.model === "bria/fibo-generate-1.5/text-to-image") {
293
+ const av = (n.fields.aspect != null && n.fields.aspect !== "") ? n.fields.aspect : "1:1";
294
+ if (av !== "") e.aspect_ratio = String(av);
295
+ }
240
296
  return e;
241
297
  }
242
298
 
@@ -255,12 +311,17 @@ function videoDims(n, ctx) {
255
311
  const aWire = p.aspect_ratio ? "aspect_ratio" : p.orientation ? "orientation" : p.resolution_ratio ? "resolution_ratio" : "aspect_ratio";
256
312
  const dP = p.duration || p.seconds;
257
313
  const dWire = p.duration ? "duration" : p.seconds ? "seconds" : "duration";
314
+ // tvideo/ivideo soft-send aspect+duration (those nodes always show the knobs).
315
+ // vedit/lipsync and resolution never forward a leftover the model did not list.
316
+ // Gallery: Omni v1 has duration+aspect only; leftover 720p/4k from v1.1 must not POST.
317
+ // Matches play videoDimParams (pack.resolution / dimDefs) — including catalog-miss omit.
318
+ const soft = n.type !== "vedit" && n.type !== "lipsync";
258
319
  let asp = f.aspect, dur = f.duration;
259
320
  if ((asp == null || asp === "") && aP && aP.default != null) asp = aP.default;
260
321
  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;
322
+ if (p.resolution && f.resolution != null && f.resolution !== "") out.resolution = f.resolution;
323
+ if ((aP || soft) && asp != null && asp !== "") out[aWire] = asp;
324
+ if ((dP || soft) && dur != null && dur !== "") out[dWire] = dur;
264
325
  return out;
265
326
  }
266
327
 
@@ -268,10 +329,13 @@ function videoDims(n, ctx) {
268
329
  Video models disagree on the ref-array param name AND its size limit; sending the wrong
269
330
  key silently degrades to a plain video, sending too many can over-bill. Resolve the
270
331
  model's REAL key from the catalog and clamp to its declared max. */
271
- function refMaxFor(model) {
332
+ function refMaxFor(model, pricing) {
272
333
  const id = String(model || "");
273
334
  if (/seedance/i.test(id)) return 9;
335
+ if (/minimax-h3/i.test(id)) return 9; // upstream cap; the first 5 are included, 6-9 bill extra_reference_image each
274
336
  if (/luma|ray/i.test(id)) return 4;
337
+ const inc = pricing && +pricing.included_reference_images;
338
+ if (inc > 0) return inc; // no family entry: the included count is the only number the catalog gives us
275
339
  return 4;
276
340
  }
277
341
 
@@ -285,15 +349,15 @@ function modelRefSpec(model, ctx) {
285
349
  const m = catItem(ctx && ctx.catalog, "video", model);
286
350
  if (!m) return { key: "reference_images", cap: refMaxFor(model) };
287
351
  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
352
+ const key = keys.find((k) => k in pp) || (pricingAdvertisesRefs(m.pricing) ? "reference_images" : null);
353
+ if (!key) return null; // known model with no ref-image param and no ref pricing
290
354
  const d = pp[key];
291
355
  let cap = null;
292
356
  if (d && typeof d === "object") {
293
357
  const mx = d.max != null ? d.max : d.maxItems != null ? d.maxItems : d.max_items;
294
358
  if (mx != null && +mx > 0) cap = +mx;
295
359
  }
296
- return { key, cap: cap != null ? cap : refMaxFor(model) };
360
+ return { key, cap: cap != null ? cap : refMaxFor(model, m.pricing) };
297
361
  }
298
362
 
299
363
  /** 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 +401,42 @@ function audioParams(n, ctx) {
337
401
  const durOk = !sp || (sp.min_duration != null && sp.max_duration != null); // cat:duration
338
402
  const secDurOk = !sp || (sp.min_duration != null && sp.max_duration != null && +((m.pricing || {}).per_second) > 0); // cat:secduration
339
403
  const voiceOk = !sp || (Array.isArray(sp.voices) && sp.voices.length > 0); // cat:voices
404
+ // Yue2 catalogs the file type as output_format (mp3/wav/flac). The OpenAI response_format
405
+ // key is ignored, so a wav/flac choice came back as mp3. No output_format in the catalog
406
+ // (or no catalog) keeps the old response_format send.
407
+ const putFormat = (fmt) => {
408
+ if (!nonEmpty(fmt)) return;
409
+ const of = sp && sp.output_format;
410
+ const values = of && Array.isArray(of.values) ? of.values : null;
411
+ if (values) {
412
+ if (values.indexOf(fmt) < 0) return;
413
+ const def = of.default != null ? of.default : "mp3";
414
+ if (fmt === def) return;
415
+ body.output_format = fmt;
416
+ return;
417
+ }
418
+ if (fmt !== "mp3") body.response_format = fmt;
419
+ };
340
420
  if (n.type === "music") {
341
421
  if (nonEmpty(f.lyrics)) body.lyrics = f.lyrics;
342
422
  if (f.instrumental === true || f.instrumental === "true") body.instrumental = true;
343
423
  if (nonEmpty(f.duration) && num(f.duration) != null && durOk) body.duration = num(f.duration);
344
424
  if (nonEmpty(f.negative_prompt)) body.negative_prompt = f.negative_prompt;
345
425
  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;
426
+ putFormat(f.response_format);
347
427
  } else if (n.type === "tts") {
348
428
  if (nonEmpty(f.voice) && voiceOk) body.voice = f.voice;
349
429
  if (nonEmpty(f.speed) && num(f.speed) != null && num(f.speed) !== 1) body.speed = num(f.speed); // omit when 1
350
430
  if (nonEmpty(f.instructions)) body.instructions = f.instructions;
351
- if (nonEmpty(f.response_format) && f.response_format !== "mp3") body.response_format = f.response_format;
431
+ putFormat(f.response_format);
352
432
  } else if (n.type === "remix") {
353
433
  if (nonEmpty(f.lyrics)) body.lyrics = f.lyrics;
354
434
  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;
435
+ // Format is a remix knob only when the catalog lists output_format (Yue2 music-to-music).
436
+ // A cover model with no such key must not receive a stale response_format.
437
+ const of = sp && sp.output_format;
438
+ const fmtListed = !!(of && Array.isArray(of.values) && of.values.length);
439
+ if (!sp || fmtListed) putFormat(f.response_format);
356
440
  }
357
441
  if ((f.extraJson || "").trim()) {
358
442
  try { Object.assign(body, JSON.parse(f.extraJson)); }
@@ -422,6 +506,13 @@ export const RUNNERS = {
422
506
  }
423
507
  return { video: n.fields.video };
424
508
  },
509
+ async mupload(n) {
510
+ if (!n.fields.model) {
511
+ if (optionalNode(n)) return { model: "" };
512
+ throw new NanoodleError("no 3D file — drop a .glb first");
513
+ }
514
+ return { model: n.fields.model };
515
+ },
425
516
 
426
517
  async choice(n) {
427
518
  const opts = String(n.fields.options || "").split("\n").map((s) => s.trim()).filter(Boolean);
@@ -504,7 +595,16 @@ export const RUNNERS = {
504
595
 
505
596
  async llm(n, inp, ctx) {
506
597
  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")));
598
+ // Wired image ports: a text-only chat model can't see the pictures and NanoGPT still
599
+ // bills the image tokens, so drop them for a KNOWN text-only model and note it —
600
+ // mirrors play's chatModelCan("vision") / editor modelSupportsImages gate (permissive
601
+ // for catalog-absent models; twin of the audio gate below).
602
+ const allImgs = collectPorts(inp, IMG_PORT_RE);
603
+ const seesImages = chatModelCan(ctx.catalog, mdl(n), "vision");
604
+ const imgs = seesImages
605
+ ? await Promise.all(allImgs.map((u) => fitImage(u, ctx, "wired image")))
606
+ : [];
607
+ if (allImgs.length && !seesImages) ctx.progress("image(s) ignored — this model is text-only");
508
608
  // hosted audio (music/tts nodes return https CDN URLs verbatim) → download + inline as base64:
509
609
  // the chat input_audio part carries bytes, never a URL
510
610
  let audioPart = null;
@@ -536,6 +636,17 @@ export const RUNNERS = {
536
636
  return { text: await ctx.chat(messages, mdl(n), {}) };
537
637
  },
538
638
 
639
+ // ⚖️ Decide: one typed question to a NanoGPT decision model (twin of the editor/play node).
640
+ // A closed yes/no gate throws a NanoodleError with code "decide-gate" — downstream skips, unbilled.
641
+ async decide(n, inp, ctx) {
642
+ const model = mdl(n);
643
+ const it = catItem(ctx.catalog, "chat", model);
644
+ // a catalog row WITHOUT decision_input (e.g. the non-detailed /api/v1/models list) is "unknown", not text-only
645
+ const lim = decideImageLimits(it && it.decision_input, !!(it && it.decision_input));
646
+ const fit = (u, maxDim, budget) => fitImageJpeg(u, maxDim, budget, mediaOpts(ctx));
647
+ return decideRun(n.fields || {}, model, inp.text, collectPorts(inp, IMG_PORT_RE), lim, (body) => ctx.decide(body), fit);
648
+ },
649
+
539
650
  async image(n, inp, ctx) {
540
651
  const prompt = promptOf(n, inp, "no prompt");
541
652
  let want = Math.max(1, parseInt(n.fields.variations, 10) || 1);
@@ -549,6 +660,7 @@ export const RUNNERS = {
549
660
  },
550
661
 
551
662
  async edit(n, inp, ctx) {
663
+ guardInputRoles(n.fields.model, inp); // role models: refuse holes before spending
552
664
  let imgs = collectPorts(inp, EDIT_IMG_RE);
553
665
  if (!imgs.length) throw new NanoodleError("no image input");
554
666
  // cap to the model's max_input_images (item present but silent → 1; absent → no cap):
@@ -567,7 +679,8 @@ export const RUNNERS = {
567
679
  imgs = await Promise.all(imgs.map((u) => fitImage(u, ctx, "source image")));
568
680
  guardRefsSize(imgs);
569
681
  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) }) };
682
+ const urls = await ctx.image({ prompt, model: mdl(n), size: n.fields.size || "1024x1024", imageDataUrl: src, extra: imgExtra(n), multi: true });
683
+ return { image: keptFirst(urls, ctx) };
571
684
  },
572
685
 
573
686
  async inpaint(n, inp, ctx) {
@@ -580,7 +693,8 @@ export const RUNNERS = {
580
693
  // ctx.maskToSource lets a browser host inject its canvas compositor (handles JPEG/WebP
581
694
  // sources the pure-PNG path can't, where ffmpeg isn't an option).
582
695
  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) }) };
696
+ const urls = await ctx.image({ prompt, model: mdl(n), size: n.fields.size || "1024x1024", imageDataUrl: source, maskDataUrl: mask, extra: imgExtra(n), multi: true });
697
+ return { image: keptFirst(urls, ctx) };
584
698
  },
585
699
 
586
700
  async tvideo(n, inp, ctx) {
@@ -607,7 +721,12 @@ export const RUNNERS = {
607
721
  },
608
722
 
609
723
  async lipsync(n, inp, ctx) {
610
- if (!inp.image) throw new NanoodleError("no image input");
724
+ // Portrait required only for image_to_video avatars. Preset-only talking avatars (audio→video,
725
+ // catalog image_to_video false — e.g. HeyGen Avatar V) run from the preset + audio alone, and a
726
+ // wired image is NOT forwarded. Catalog miss / no catalog keeps the portrait requirement.
727
+ // Twin of play's built-in lipsync runner (nanoodle #733).
728
+ const needsPortrait = avatarNeedsPortrait(ctx && ctx.catalog, String((n.fields && n.fields.model) || "").trim());
729
+ if (needsPortrait && !inp.image) throw new NanoodleError("no image input");
611
730
  if (!inp.audio) throw new NanoodleError("no audio input");
612
731
  const prompt = promptOf(n, inp);
613
732
  // Avatar models cap audio length (LongCat = 30s) and the cap isn't reliably in the catalog,
@@ -619,7 +738,7 @@ export const RUNNERS = {
619
738
  // ctx.trimAudio lets a browser host inject its Web Audio trimmer so the retry bytes match
620
739
  // its built-in runner exactly; default is the local-media trimmer (pure-JS WAV, ffmpeg).
621
740
  const trim = ctx.trimAudio || ((url, start, len, rate) => trimAudioToWav(url, start, len, rate, mediaOpts(ctx)));
622
- const img = await fitImage(inp.image, ctx, "portrait image");
741
+ const img = needsPortrait ? await fitImage(inp.image, ctx, "portrait image") : null;
623
742
  let trimSec = null;
624
743
  for (let attempt = 0; attempt < 2; attempt++) {
625
744
  let opts;
@@ -688,4 +807,69 @@ export const RUNNERS = {
688
807
  if (!inp.audio) throw new NanoodleError("no audio input");
689
808
  return { text: await ctx.transcribe(mdl(n), inp.audio, (n.fields.language || "auto").trim()) };
690
809
  },
810
+
811
+ // 🎧 Clean voice: hosted audio or video → voice-only audio. data:/blob: refused before any request.
812
+ async cleanvoice(n, inp, ctx) {
813
+ const model = String((n.fields && n.fields.model) || "").trim() || CLEANVOICE_DEFAULT_MODEL;
814
+ return cleanVoiceRun(model, inp, n.fields, (m, extra) => ctx.audio(m, "", extra));
815
+ },
816
+
817
+ // 🧊 3D model: image and/or text → GLB via the video submit/poll API (mediaKind model3d).
818
+ async model3d(n, inp, ctx) {
819
+ const model = String((n.fields && n.fields.model) || "").trim() || MODEL3D_IMAGE_DEFAULT;
820
+ const mods = model3dInputMods(ctx && ctx.catalog, model);
821
+ const prompt = promptOf(n, inp);
822
+ let image = mods.image ? (inp.image || "") : "";
823
+ if (inp.image && !mods.image && ctx && ctx.progress) {
824
+ ctx.progress("This model takes text only — the image wire stays, but this run ignores it.");
825
+ }
826
+ if (prompt && !mods.text && ctx && ctx.progress) {
827
+ ctx.progress("This model takes an image only — the prompt stays, but this run ignores it.");
828
+ }
829
+ if (mods.image && !mods.text && !image) throw new NanoodleError("Connect an image first.");
830
+ if (mods.text && !mods.image && !prompt) throw new NanoodleError("Add a prompt first.");
831
+ if (mods.image && mods.text && !image && !prompt) throw new NanoodleError("Add a prompt or connect an image first.");
832
+ if (image) image = await fitImage(image, ctx, "source image");
833
+ const url = await ctx.video(model, mods.text ? prompt : "", {
834
+ extra: n.fields.modelOpts || {},
835
+ mediaKind: "model3d",
836
+ }, image || null);
837
+ return { model: url };
838
+ },
839
+
840
+ // 🔌 Custom endpoint: POST to the graph's own URL. ctx.fetch so tests can point it at a mock.
841
+ async endpoint(n, inp, ctx) {
842
+ return runEndpoint(n, inp, { fetch: ctx && ctx.fetch, signal: ctx && ctx.signal });
843
+ },
691
844
  };
845
+
846
+ /**
847
+ * Does this avatar/lipsync model need a portrait image? True unless the model is in the
848
+ * opt-in catalog and is preset-only (no image_to_video — audio→video talking avatars such as
849
+ * HeyGen Avatar V). Catalog miss → true (the long-standing portrait requirement).
850
+ * Mirrors play's built-in lipsync runner: raw `capabilities.image_to_video` or a normalized `i2v`.
851
+ */
852
+ function avatarNeedsPortrait(catalog, model) {
853
+ const m = catItem(catalog, "video", model);
854
+ return !m || !!((m.capabilities || {}).image_to_video) || !!m.i2v;
855
+ }
856
+
857
+ // Image→3D (Tripo) takes a photo only. Hunyuan Rapid takes text and an optional image.
858
+ // Anything else, including a catalog miss, accepts either — the editor's same fallback.
859
+ const MODEL3D_IMAGE_DEFAULT = "tripo3d/v2.5";
860
+ const MODEL3D_TEXT_DEFAULT = "wavespeed-ai/hunyuan-3d-v3.1-rapid";
861
+
862
+ function model3dInputMods(catalog, id) {
863
+ const it = catItem(catalog, "model3d", id);
864
+ const raw = it && (
865
+ (Array.isArray(it.modalities) && it.modalities)
866
+ || (it.architecture && Array.isArray(it.architecture.input_modalities) && it.architecture.input_modalities)
867
+ );
868
+ if (raw && raw.length) {
869
+ const mods = raw.filter((x) => x === "image" || x === "text");
870
+ if (mods.length) return { image: mods.includes("image"), text: mods.includes("text") };
871
+ }
872
+ if (id === MODEL3D_IMAGE_DEFAULT) return { image: true, text: false };
873
+ if (id === MODEL3D_TEXT_DEFAULT) return { image: true, text: true };
874
+ return { image: true, text: true };
875
+ }