Upload folder using huggingface_hub

README.md

@@ -1,3 +1,334 @@
 ---
-license: apache-2.0
+base_model: Qwen/Qwen3.6-27B
+library_name: peft
+tags:
+  - peft
+  - lora
+  - prompt-rewriting
+  - minimax-h3
+  - text-to-audio-video
+  - audio-video-generation
+  - t2va
+language:
+  - en
 ---
+
+# MiniMax-H3 T2VA Prompt Rewriter LoRA
+
+[![LightX2V](https://img.shields.io/badge/Inference-LightX2V-ff6b35)](https://github.com/ModelTC/LightX2V)
+[![GitHub stars](https://img.shields.io/github/stars/ModelTC/LightX2V?style=social)](https://github.com/ModelTC/LightX2V)
+[![MiniMax-H3](https://img.shields.io/badge/Base%20Generator-MiniMax--H3-yellow)](https://huggingface.co/MiniMaxAI/MiniMax-H3)
+
+An open, local prompt rewriter for **text-to-audio-video (T2VA)** generation with
+[MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3), fine-tuned as a LoRA
+adapter on top of [Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B).
+
+> **Generate the final audio-video with
+> [LightX2V](https://github.com/ModelTC/LightX2V).** LightX2V provides optimized local
+> MiniMax-H3 inference, including T2AV and other H3 tasks, memory offloading, tensor and
+> sequence parallelism, quantized checkpoints, optimized attention/RMSNorm/RoPE kernels,
+> and feature caching. If this project is useful, please consider giving
+> **[ModelTC/LightX2V](https://github.com/ModelTC/LightX2V) a star**.
+
+The complete local workflow is:
+
+```text
+Short prompt ──► this Prompt Rewriter LoRA ──► structured H3 prompt
+                                                        │
+Official MiniMax-H3 weights ──► LightX2V inference ◄─────┘
+                                      │
+                                      ▼
+                          synchronized video + audio
+```
+
+The [official MiniMax-H3 repository](https://huggingface.co/MiniMaxAI/MiniMax-H3)
+provides the upstream model weights. **LightX2V is the recommended inference entry
+point for the workflow documented here.**
+
+This is an unofficial, community-trained project and is not affiliated with or
+endorsed by MiniMax. It is intended as a local alternative for producing structured,
+H3-oriented T2VA prompts when the hosted H3-Context-IR service is unavailable or
+unsuitable for a workflow. It does not reproduce the proprietary H3-Context-IR system,
+and exact output parity is not guaranteed.
+
+## What it does
+
+The adapter turns a short prompt plus the requested aspect ratio and duration into a
+structured audio-video description:
+
+```text
+Original prompt + aspect ratio + duration
+                    │
+                    ▼
+          Qwen3.6-27B + this LoRA
+                    │
+                    ▼
+integrated_multimodal_description: [Shot 1] ...
+overall_soundscape: ...
+non_diegetic_music: ...
+```
+
+The rewritten prompt expands shot structure, timing, composition, camera motion,
+physical action, continuity, synchronized diegetic sound, and non-diegetic music while
+preserving the original intent.
+
+| Component | Purpose | Model |
+|---|---|---|
+| Inference framework | Runs MiniMax-H3 locally and generates the final synchronized video and audio | **[ModelTC/LightX2V](https://github.com/ModelTC/LightX2V)** |
+| Generator weights | Upstream MiniMax-H3 model weights consumed by LightX2V | [MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) |
+| Base language model | General instruction-following backbone | [Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B) |
+| LoRA adapter | H3-oriented T2VA prompt rewriting | [lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA](https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA) |
+
+## Status and roadmap
+
+- [x] T2VA prompt rewriting
+- [x] Duration conditioning from 4 to 15 seconds
+- [x] Aspect-ratio conditioning: `21:9`, `16:9`, `4:3`, `1:1`, `3:4`, `9:16`
+- [x] Single-prompt inference script
+- [x] End-to-end MiniMax-H3 generation through [LightX2V](https://github.com/ModelTC/LightX2V)
+- [ ] First/last-frame-to-audio-video (FL2VA) prompt rewriting
+- [ ] Reference-to-audio-video (Ref2VA) prompt rewriting
+- [ ] Image-to-audio-video and last-frame variants
+- [ ] Image, video, and audio reference inputs
+- [ ] Quantitative evaluation and a reproducible public benchmark suite
+- [ ] Smaller and quantized rewriter variants
+
+Only **T2VA** is supported by this release. The remaining MiniMax-H3 tasks require
+multimodal inputs and task-specific training data; they will be added in future work.
+
+## Installation
+
+Clone this repository and install the inference dependencies:
+
+```bash
+git clone https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA
+cd MiniMax-H3-Prompt-Rewriter-LoRA
+pip install -r requirements.txt
+```
+
+The 27B base model is not included in this adapter repository and will be downloaded
+from Hugging Face on first use. Ensure that the machine has enough GPU/CPU memory for
+Qwen3.6-27B. `device_map="auto"` is used to distribute the model over available devices.
+
+## Rewrite one prompt
+
+```bash
+python infer.py \
+  --prompt "Epic space-opera teaser: a captain watches the last fleet jump away, leaving her alone." \
+  --duration 10 \
+  --resolution 16:9 \
+  --greedy
+```
+
+The default model IDs are:
+
+```text
+Base model: Qwen/Qwen3.6-27B
+Adapter:    lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA
+```
+
+Local paths are supported as well:
+
+```bash
+python infer.py \
+  --base-model /path/to/Qwen3.6-27B \
+  --adapter /path/to/MiniMax-H3-Prompt-Rewriter-LoRA \
+  --prompt "A red fox walks through a snowy forest at dawn." \
+  --duration 15 \
+  --resolution 21:9 \
+  --output rewritten_prompt.json
+```
+
+Use `--output result.txt` for plain text or a `.json` path for a record containing
+the original prompt, conditions, model IDs, and rewritten prompt.
+
+To run the base-model baseline without the LoRA:
+
+```bash
+python infer.py \
+  --base-only \
+  --prompt "A red fox walks through a snowy forest at dawn." \
+  --duration 15 \
+  --resolution 21:9 \
+  --greedy
+```
+
+For fair comparisons, use the same decoding mode and seed for every rewriter. Greedy
+decoding is the simplest deterministic setting; sampling is enabled when `--greedy`
+is omitted.
+
+## Output format
+
+The model is trained to return exactly three fields in this order:
+
+```text
+integrated_multimodal_description: [Shot 1] ... [Shot 2] At 00:05.000, ...
+overall_soundscape: ...
+non_diegetic_music: ...
+```
+
+Pass the entire three-field result to the MiniMax-H3 T2VA pipeline in
+[LightX2V](https://github.com/ModelTC/LightX2V).
+
+## Generate audio-video with LightX2V
+
+### 1. Install LightX2V
+
+Building from source gives access to the latest MiniMax-H3 integration:
+
+```bash
+git clone https://github.com/ModelTC/LightX2V.git
+cd LightX2V
+uv pip install -v .  # or: pip install -v .
+```
+
+Follow the [LightX2V Quick Start](https://lightx2v-en.readthedocs.io/en/latest/getting_started/quickstart.html)
+to install the optional optimized attention and quantization operators required by the
+selected configuration.
+
+### 2. Download the MiniMax-H3 weights
+
+The generator weights come from the upstream
+[MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) repository:
+
+```bash
+hf download MiniMaxAI/MiniMax-H3 --local-dir /path/to/MiniMax-H3
+```
+
+The LoRA in this repository rewrites text; it does not replace or include the
+MiniMax-H3 generator weights.
+
+### 3. Rewrite the prompt
+
+Run this command from the prompt-rewriter repository:
+
+```bash
+python infer.py \
+  --prompt "A red fox walks through a snowy forest at dawn." \
+  --duration 15 \
+  --resolution 16:9 \
+  --greedy \
+  --output /tmp/h3_rewritten_prompt.txt
+```
+
+### 4. Generate synchronized video and audio
+
+Run the rewritten prompt with LightX2V from the `LightX2V` repository root:
+
+```bash
+source scripts/base/base.sh
+export DTYPE=BF16
+export SENSITIVE_LAYER_DTYPE=FP32
+
+python -m lightx2v.infer \
+  --model_cls minimax_h3 \
+  --task t2av \
+  --model_path /path/to/MiniMax-H3 \
+  --config_json configs/minimax_h3/minimax_h3_t2av.json \
+  --prompt "$(</tmp/h3_rewritten_prompt.txt)" \
+  --save_result_path save_results/minimax_h3_t2av.mp4 \
+  --seed 42
+```
+
+The output MP4 contains both video and synchronized stereo audio. Make sure that the
+duration and aspect ratio in the LightX2V config match the conditions passed to the
+prompt rewriter. Ready-to-run single-GPU, block-offload, tensor-parallel, and combined
+TP+SP examples are available in the
+**[LightX2V MiniMax-H3 scripts](https://github.com/ModelTC/LightX2V/tree/main/scripts/minimax_h3)**.
+
+For lower-memory or multi-GPU inference, start from one of these configurations:
+
+| Mode | LightX2V example |
+|---|---|
+| Single GPU | `configs/minimax_h3/minimax_h3_t2av.json` |
+| Block offload | `configs/minimax_h3/minimax_h3_t2av_block_offload.json` |
+| Sequence parallelism | `configs/minimax_h3/minimax_h3_t2av_sp.json` |
+| Tensor parallelism | `configs/minimax_h3/minimax_h3_t2av_tp.json` |
+| Tensor + sequence parallelism | `configs/minimax_h3/minimax_h3_t2av_tp_sp.json` |
+
+## Qualitative comparison
+
+The following gallery is designed to compare videos generated through
+[LightX2V](https://github.com/ModelTC/LightX2V) with the same MiniMax-H3 checkpoint and
+identical inference settings. Only the prompt-rewriting method should differ.
+
+| Label | Meaning |
+|---|---|
+| Original Prompt (No Rewrite) | The raw user prompt is sent directly to MiniMax-H3 |
+| MiniMax Context-IR | The raw prompt is enhanced by the official hosted Context-IR workflow |
+| Qwen3.6-27B Base Rewriter | The same template is used with the base model and no LoRA |
+| **H3-T2VA Context Rewriter LoRA (Ours)** | Qwen3.6-27B with this task-specific LoRA |
+
+### Prompt 1
+
+> **Original prompt:** TODO — paste Prompt 1 here  
+> **Conditions:** TODO — duration, aspect ratio, seed, and generation configuration
+
+| Original Prompt (No Rewrite) | MiniMax Context-IR | Qwen3.6-27B Base Rewriter | **H3-T2VA Context Rewriter LoRA (Ours)** |
+|---|---|---|---|
+| <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt1/original.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt1/minimax_context_ir.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt1/qwen_base.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt1/ours.mp4"></video> |
+
+### Prompt 2
+
+> **Original prompt:** TODO — paste Prompt 2 here  
+> **Conditions:** TODO — duration, aspect ratio, seed, and generation configuration
+
+| Original Prompt (No Rewrite) | MiniMax Context-IR | Qwen3.6-27B Base Rewriter | **H3-T2VA Context Rewriter LoRA (Ours)** |
+|---|---|---|---|
+| <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt2/original.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt2/minimax_context_ir.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt2/qwen_base.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt2/ours.mp4"></video> |
+
+### Prompt 3
+
+> **Original prompt:** TODO — paste Prompt 3 here  
+> **Conditions:** TODO — duration, aspect ratio, seed, and generation configuration
+
+| Original Prompt (No Rewrite) | MiniMax Context-IR | Qwen3.6-27B Base Rewriter | **H3-T2VA Context Rewriter LoRA (Ours)** |
+|---|---|---|---|
+| <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt3/original.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt3/minimax_context_ir.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt3/qwen_base.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt3/ours.mp4"></video> |
+
+### Prompt 4
+
+> **Original prompt:** TODO — paste Prompt 4 here  
+> **Conditions:** TODO — duration, aspect ratio, seed, and generation configuration
+
+| Original Prompt (No Rewrite) | MiniMax Context-IR | Qwen3.6-27B Base Rewriter | **H3-T2VA Context Rewriter LoRA (Ours)** |
+|---|---|---|---|
+| <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt4/original.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt4/minimax_context_ir.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt4/qwen_base.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt4/ours.mp4"></video> |
+
+### Prompt 5
+
+> **Original prompt:** TODO — paste Prompt 5 here  
+> **Conditions:** TODO — duration, aspect ratio, seed, and generation configuration
+
+| Original Prompt (No Rewrite) | MiniMax Context-IR | Qwen3.6-27B Base Rewriter | **H3-T2VA Context Rewriter LoRA (Ours)** |
+|---|---|---|---|
+| <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt5/original.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt5/minimax_context_ir.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt5/qwen_base.mp4"></video> | <video controls src="https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA/resolve/main/assets/examples/prompt5/ours.mp4"></video> |
+
+Expected video paths and filenames are documented in
+[`assets/examples/UPLOAD_LAYOUT.md`](assets/examples/UPLOAD_LAYOUT.md).
+
+## Limitations
+
+- This release supports text-only T2VA prompt rewriting; it does not consume images,
+  videos, or audio references.
+- It is a learned approximation, not an open-source release or exact replica of the
+  official H3-Context-IR service.
+- The generated rewrite may add details not explicitly present in a short prompt.
+  Review the result before generation when identity, dialogue, timing, or composition
+  must be exact.
+- Final quality depends on both prompt rewriting and the MiniMax-H3 inference settings
+  selected in LightX2V.
+- Use of MiniMax-H3 is governed by the license and acceptable-use terms published in
+  the [official MiniMax-H3 repository](https://huggingface.co/MiniMaxAI/MiniMax-H3).
+
+## Acknowledgements
+
+This project is developed by the
+[LightX2V](https://github.com/ModelTC/LightX2V) community to make local MiniMax-H3
+prompt rewriting and audio-video generation easier and more efficient. It is built for
+the [MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) audio-video generation
+model and fine-tuned from [Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B).
+
+If you use this adapter or the MiniMax-H3 integration, please visit and star
+**[ModelTC/LightX2V](https://github.com/ModelTC/LightX2V)**, where inference support,
+performance optimizations, examples, and future task updates are maintained.

infer.py

@@ -0,0 +1,143 @@
+#!/usr/bin/env python3
+"""Rewrite one T2VA prompt with the MiniMax-H3 Prompt Rewriter LoRA."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+import transformers
+from transformers import AutoTokenizer, set_seed
+
+from prompt_template import build_messages
+
+
+DEFAULT_BASE_MODEL = "Qwen/Qwen3.6-27B"
+DEFAULT_ADAPTER = "lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA"
+RESOLUTIONS = ("21:9", "16:9", "4:3", "1:1", "3:4", "9:16")
+
+
+def get_model_class():
+    model_class = getattr(transformers, "AutoModelForImageTextToText", None)
+    if model_class is None:
+        model_class = getattr(transformers, "AutoModelForVision2Seq", None)
+    if model_class is None:
+        raise RuntimeError(
+            "A recent Transformers version with AutoModelForImageTextToText "
+            "support is required. Install the packages in requirements.txt."
+        )
+    return model_class
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--prompt", required=True, help="Original text prompt to rewrite.")
+    parser.add_argument("--duration", type=int, choices=range(4, 16), default=10, metavar="4..15")
+    parser.add_argument("--resolution", choices=RESOLUTIONS, default="16:9", help="Target aspect ratio.")
+    parser.add_argument("--base-model", default=DEFAULT_BASE_MODEL, help="HF model ID or local base-model path.")
+    parser.add_argument("--adapter", default=DEFAULT_ADAPTER, help="HF model ID or local PEFT adapter path.")
+    parser.add_argument(
+        "--base-only",
+        action="store_true",
+        help="Skip the LoRA and run the Qwen base-model baseline.",
+    )
+    parser.add_argument("--max-new-tokens", type=int, default=2048)
+    parser.add_argument("--dtype", choices=("bfloat16", "float16", "float32"), default="bfloat16")
+    parser.add_argument("--attn-implementation", choices=("sdpa", "flash_attention_2", "eager"), default="sdpa")
+    parser.add_argument("--temperature", type=float, default=0.7)
+    parser.add_argument("--top-p", type=float, default=0.8)
+    parser.add_argument("--top-k", type=int, default=20)
+    parser.add_argument("--repetition-penalty", type=float, default=1.05)
+    parser.add_argument("--greedy", action="store_true", help="Use deterministic greedy decoding.")
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument(
+        "--output",
+        type=Path,
+        help="Optional output path. A .json file stores conditions and output; other suffixes store plain text.",
+    )
+    return parser.parse_args()
+
+
+def input_device(model: torch.nn.Module) -> torch.device:
+    """Return the embedding device, including when Accelerate shards the model."""
+    embeddings = model.get_input_embeddings()
+    if embeddings is not None and hasattr(embeddings, "weight"):
+        return embeddings.weight.device
+    return next(model.parameters()).device
+
+
+def main() -> None:
+    args = parse_args()
+    set_seed(args.seed)
+    dtype = getattr(torch, args.dtype)
+
+    tokenizer = AutoTokenizer.from_pretrained(args.base_model, trust_remote_code=True)
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    model = get_model_class().from_pretrained(
+        args.base_model,
+        torch_dtype=dtype,
+        device_map="auto",
+        low_cpu_mem_usage=True,
+        trust_remote_code=True,
+        attn_implementation=args.attn_implementation,
+    )
+
+    if not args.base_only:
+        from peft import PeftModel
+
+        model = PeftModel.from_pretrained(model, args.adapter)
+    model.eval()
+
+    rendered = tokenizer.apply_chat_template(
+        build_messages(args.prompt, args.resolution, args.duration),
+        tokenize=False,
+        add_generation_prompt=True,
+        enable_thinking=False,
+    )
+    inputs = tokenizer(rendered, return_tensors="pt", add_special_tokens=False)
+    device = input_device(model)
+    inputs = {name: tensor.to(device) for name, tensor in inputs.items()}
+
+    generation_kwargs = {
+        "max_new_tokens": args.max_new_tokens,
+        "do_sample": not args.greedy,
+        "repetition_penalty": args.repetition_penalty,
+        "pad_token_id": tokenizer.pad_token_id,
+        "eos_token_id": tokenizer.eos_token_id,
+    }
+    if not args.greedy:
+        generation_kwargs.update(
+            temperature=args.temperature,
+            top_p=args.top_p,
+            top_k=args.top_k,
+        )
+
+    with torch.inference_mode():
+        generated = model.generate(**inputs, **generation_kwargs)
+    new_tokens = generated[:, inputs["input_ids"].shape[1] :]
+    rewritten_prompt = tokenizer.batch_decode(new_tokens, skip_special_tokens=True)[0].strip()
+
+    if args.output is not None:
+        args.output.parent.mkdir(parents=True, exist_ok=True)
+        if args.output.suffix.lower() == ".json":
+            record = {
+                "prompt": args.prompt.strip(),
+                "resolution": args.resolution,
+                "duration": args.duration,
+                "rewritten_prompt": rewritten_prompt,
+                "base_model": args.base_model,
+                "adapter": None if args.base_only else args.adapter,
+            }
+            args.output.write_text(json.dumps(record, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+        else:
+            args.output.write_text(rewritten_prompt + "\n", encoding="utf-8")
+
+    print(rewritten_prompt)
+
+
+if __name__ == "__main__":
+    main()

prompt_template.py

@@ -0,0 +1,40 @@
+"""Prompt template shared by MiniMax-H3 T2VA prompt-rewriter inference."""
+
+from __future__ import annotations
+
+
+SYSTEM_PROMPT = """You are a professional prompt rewriter for joint audio-video generation.
+Rewrite the user's original prompt into one coherent, production-ready multimodal description for the requested output aspect ratio and duration.
+
+Return only these three fields, in this exact order:
+integrated_multimodal_description: ...
+overall_soundscape: ...
+non_diegetic_music: ...
+
+Requirements:
+- Expand the visual narrative into clearly numbered shots such as [Shot 1], [Shot 2], and include timestamps for cuts after the first shot when useful.
+- Make the number, timing, and pacing of shots appropriate for the requested duration.
+- Compose the scene for the requested aspect ratio.
+- Preserve the user's intent while adding concrete subjects, appearance, environment, lighting, composition, camera movement, physical motion, and temporal continuity.
+- Keep characters, objects, wardrobe, locations, and spatial relationships consistent across shots.
+- Describe synchronized diegetic audio in overall_soundscape and external score in non_diegetic_music.
+- Do not add explanations, Markdown fences, safety commentary, or fields other than the three requested fields."""
+
+
+def build_messages(prompt: str, resolution: str, duration: int) -> list[dict[str, str]]:
+    """Build the chat messages used during LoRA training and inference."""
+    prompt = prompt.strip()
+    if not prompt:
+        raise ValueError("prompt must not be empty")
+
+    return [
+        {"role": "system", "content": SYSTEM_PROMPT},
+        {
+            "role": "user",
+            "content": (
+                f"resolution: {resolution}\n"
+                f"duration: {duration}s\n"
+                f"original_prompt: {prompt}"
+            ),
+        },
+    ]

requirements.txt

@@ -0,0 +1,5 @@
+torch>=2.6
+transformers>=4.57.1
+accelerate>=1.10
+peft>=0.17
+safetensors>=0.5