# AGENTS.md — ComfyUI-Auto_installer-Assets

## Purpose

Pure assets repository (HuggingFace LFS + ModelScope mirror) for model files and installation binaries used by [ComfyUI-Auto_installer](https://github.com/UmeAiRT/ComfyUI-Auto_installer). No code — only large binary files.

## Ecosystem

This repo is consumed by:
- **[ComfyUI-Auto_installer](https://github.com/UmeAiRT/ComfyUI-Auto_installer)** — PowerShell download scripts reference these files by exact path
- **[ComfyUI-UmeAiRT-Toolkit](https://github.com/UmeAiRT/ComfyUI-UmeAiRT-Toolkit)** — Automatic model downloader nodes pull from this repo

Any file rename or move here **must** be reflected in both projects' download logic.

**Mirrors:**
- **HuggingFace:** [UmeAiRT/ComfyUI-Auto_installer](https://huggingface.co/UmeAiRT/ComfyUI-Auto_installer)
- **ModelScope:** [UmeAiRT/ComfyUI-Auto-Installer-Assets](https://www.modelscope.ai/datasets/UmeAiRT/ComfyUI-Auto-Installer-Assets)

## Folder Structure

| Folder | Contents | Notes |
|---|---|---|
| `models/diffusion_models/` | Backbone models (GGUF + safetensors) | Official ComfyUI folder. Organized by family (`FLUX/`, `WAN/`, `QWEN/`, `HiDream/`, `LTX-2/`, `LTXV/`, `Z-IMG/`, `MelBandRoFormer/`) |
| `models/text_encoders/` | Text encoders (T5, QWEN, GEMMA, LLAMA) | Official ComfyUI folder. **NOT** `clip/` |
| `models/clip/` | CLIP models only (clip_l, clip_g, ViT, longclip) | Different from `text_encoders/`. Do NOT put text encoders here |
| `models/clip_vision/` | CLIP Vision encoders (clip_vision_h, sigclip) | Used by IP-Adapter / Redux |
| `models/checkpoints/` | Full merged checkpoints | |
| `models/loras/` | LoRA adapters | Organized by family (`FLUX/`, `WAN2.1/`, `WAN2.2/`, `QWEN/`, `LTX-2/`) |
| `models/vae/` | VAE models (transverse, shared across families) | |
| `models/vae_approx/` | Approximate VAE decoders (TAESD) | For fast preview |
| `models/controlnet/` | ControlNet models | |
| `models/upscale_models/` | Upscaler models (RealESRGAN, etc.) | |
| `models/style_models/` | Style/Redux models | |
| `models/pulid/` | PuLID face models | |
| `models/sams/` | SAM segmentation models | |
| `models/ultralytics/` | YOLO detection models (face, hand) | |
| `models/xlabs/` | XLabs custom node models | ⚠️ Hardcoded path — do NOT move |
| `models/onnx/` | ONNX upscalers for TensorRT | ⚠️ Hardcoded path — do NOT move |
| `models/tensorrt/` | Pre-built TensorRT engines | ⚠️ Hardcoded path — do NOT move |
| `models/LLM/` | Language models (Florence PromptGen) | Multi-file model directories |
| `bin/` | Installation executables (Git, Python, CUDA, etc.) | |
| `whl/` | Pre-compiled Python wheels | |
| `packages/` | Zip archives (extensions, build tools, configs) | |

## Naming Conventions

### GGUF files — Follow the [official GGUF spec](https://github.com/ggml-org/ggml/blob/master/docs/gguf.md)

Format: `<BaseName>-<SizeLabel>-<FineTune>-<Encoding>.gguf`
- **CamelCase** BaseName: `Flux1`, `Wan2.1`, `HiDream`, `LTXV`, `LTX-2`, `Qwen`
- **Uppercase** SizeLabel: `14B`, `1.3B`, `13B`
- **Uppercase** Encoding: `Q4_K_S`, `Q8_0`, `Q6_K`
- Hyphens between components

Examples: `Flux1-Dev-Q4_K_S.gguf`, `Wan2.2-T2V-HighNoise-14B-Q8_0.gguf`

### Safetensors files — Community convention

- **Lowercase** with **hyphens** as separators
- Precision suffixes: `fp16`, `bf16`, `fp8_e4m3fn`, `fp8_scaled`

Examples: `wan2.1-t2v-14b-fp16.safetensors`, `qwen-image-edit-fp8_e4m3fn.safetensors`

## Quantization Coverage Target

Each diffusion model family should provide these versions when available upstream:
- **Full precision:** `bf16` or `fp16` (safetensors)
- **Reduced precision:** `fp8_e4m3fn` or `fp8_scaled` (safetensors)
- **GGUF quantizations:** `Q8_0`, `Q6_K`, `Q5_K_S`, `Q4_K_S`, `Q3_K_S`