@melite/models-music3 0.1.4-alpha → 0.1.5-pre-alpha

package/README.md

@@ -0,0 +1,45 @@
+---
+description: MiniMax Music 3 text-to-music — caption and lyrics conditioning over INT8 ConvRot weights, mp3 output under the run directory.
+kind: package-library
+---
+
+# music3
+
+music3 registers the MiniMax Music 3 family (form `form_music`, label "MiniMax Music 3", genre `audio`, engine `comfyui`): text-to-music with lyrics. Weights: three INT8 ConvRot provisions from Comfy-Org/MiniMax-Music-3 — the DiT (`minimax_music3_dit_int8_convrot.safetensors`, about 2.5 GB), the pruned text encoder (about 9.2 GB), and the DAV VAE (about 0.22 GB). The `lane_inputs` pin `clip_type: minimax` and the three file names.
+
+## The compose action
+
+`compose` builds: UNETLoader + CLIPLoader (type `minimax`) + VAELoader → MiniMaxMusic3TextEncode (caption + lyrics, `top_k` 50) → ConditioningZeroOut (the negative arm) → EmptyMiniMaxMusic3LatentAudio → KSampler (euler / simple) → VAEDecodeAudio → SaveAudioAdvanced (mp3, V0) under `<run>/music`.
+
+Params: `prompt` (required — the caption), `lyrics`, `duration` [5, 360] seconds, `guidance_scale` [0.1, 10], `lm_cfg_scale` [0.1, 10], `n_diffusion_steps` [1, 200], `seed` (seed-sentinel rule).
+
+## Variants and recipes
+
+One variant (`standard`, INT8 ConvRot cell, about 20 GB VRAM peak). Recipes: `short_sting` (duration 5), `full_length` (duration 360), and the no-op `default`.
+
+## Model Experience
+
+### Request context and condition
+
+A run needs the caption; lyrics, duration, and both CFG scales ride defaults. Use `short_sting` or `full_length` to jump to the duration bounds, or state `duration` directly.
+
+#### What the model sees
+
+The caption and lyrics, the duration, both CFG scales, and the step count; the solver pair (euler / simple) is fixed by the build, not a param.
+
+#### Token effect
+
+Zero-direct: an engine graph, not a token stream.
+
+#### KV Cache effect
+
+Independent: each run is self-contained.
+
+## Known Limitations and Deferred Work
+
+- **Fixed solver** — sampler and scheduler are hardcoded in the build; the params surface offers no solver override.
+- **Lyrics optional by type only** — the DECL requires just the caption; lyric-less output quality is not verified against the running model.
+- **Single quant** — one INT8 ConvRot tier; no bf16 cell is declared.
+- **Provision weight** — roughly 12 GB of downloads before the first note.
+
+Exports: the plugin root, `./package.json`.

package/lib/index.d.ts

@@ -1,6 +1,7 @@
 import type { Plugin } from "@deepseek-ai/cordis";
-import type { ModelDecl } from "@melite/comfyui-cordis/model-contract";
+import type { ModelDecl } from '@melite/comfyui-card-kit/model-contract';
 export declare const DECL: ModelDecl;
 export declare const family: Plugin;
 export { compose, parseMusicFacts, type MusicFacts, type MusicParams } from "./actions.js";
+export { family as default };
 //# sourceMappingURL=index.d.ts.map

package/lib/index.js

@@ -7,7 +7,7 @@ export const DECL = {
     "engine": "comfyui",
     "source": {
         "type": "huggingface",
-        "repo_id": "audio-cpp/MiniMax-Music3-GGUF",
+        "repo_id": "Comfy-Org/MiniMax-Music-3",
         "note": ""
     },
     "genre": "audio",
@@ -19,9 +19,9 @@ export const DECL = {
             "route": "music3",
             "cells": [
                 {
-                    "source": {
-                        "filename": "MiniMax-Music3-GGUF"
-                    },
+                    
+                    // Comfy-Org safetensors the provisions rows carry, never a
+                    // GGUF tree.
                     "quant": "int8",
                     "label": "INT8 ConvRot",
                     "route": "music3",
@@ -118,6 +118,12 @@ export const DECL = {
             },
         ],
     },
+    // THE LANE SURFACE (the z-image precedent): exactly the knobs the
+    // music card serves.
+    "params": [
+        "prompt", "lyrics", "duration", "mode",
+        "guidance_scale", "n_diffusion_steps", "lm_cfg_scale", "seed",
+    ],
     // THE DECLARED ACTIONS (the atom vocabulary, stroke 3):
     // the compose action the music card binds (the MiniMax Music 3 lane).
     // The admission gate runs the facts parse at registration.
@@ -131,4 +137,9 @@ export const family = {
     },
 };
 export { compose, parseMusicFacts } from "./actions.js";
+// The Loader-row face: DSH loads plugin rows by bare specifier and
+// unwraps `exports.default ?? exports` (vendor/cordis_plugins/loader
+// unwrapExports) — the default export IS the row-loadable plugin. The
+// named `family` stays the library face (engine loadFamilies).
+export { family as default };
 //# sourceMappingURL=index.js.map

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@melite/models-music3",
-  "version": "0.1.4-alpha",
+  "version": "0.1.5-pre-alpha",
   "license": "SSPL-1.0",
   "type": "module",
   "exports": {
@@ -11,8 +11,7 @@
     "./package.json": "./package.json"
   },
   "dependencies": {
-    "@melite/comfyui-card-kit": "0.1.4-alpha",
-    "@melite/comfyui-cordis": "0.1.4-alpha"
+    "@melite/comfyui-card-kit": "0.1.5-pre-alpha"
   },
   "devDependencies": {
     "@types/node": "^22.20.0",