Skip to content

models.md

Latest commit

 

History

History
75 lines (56 loc) · 3.28 KB

models.md

File metadata and controls

75 lines (56 loc) · 3.28 KB
Raw

Models & the shared store

All weights live in ONE store outside the repo so MFLUX, ComfyUI and other tools (opencode, a second ComfyUI, Draw Things, …) reuse a single download.

Shared store layout

~/ai-models/                 ($AI_MODELS_DIR — override to relocate)
├── hf/                       $HF_HOME  — HuggingFace cache (MFLUX, diffusers, opencode)
└── comfyui/                  $COMFY_MODELS_DIR — ComfyUI single-file weights
    ├── checkpoints/  diffusion_models/  unet/
    ├── text_encoders/  clip/  clip_vision/
    ├── vae/  loras/  controlnet/  upscale_models/  embeddings/  style_models/

Two sharing mechanisms, because there are two model conventions:

  • HF-cache tools (MFLUX, diffusers, opencode) honour HF_HOME → point them at ~/ai-models/hf and it's one copy, many tools. env.sh sets this.
  • ComfyUI single-file weights are shared via comfyui/extra_model_paths.yaml (auto-written by comfyui-start.sh) pointing at ~/ai-models/comfyui.

A model used in both ecosystems may exist as two formats (HF diffusers repo vs. repacked single file). Within each ecosystem, sharing is a single copy.

Reusing across tools (opencode, etc.)

  • HF-based tools: export HF_HOME=~/ai-models/hf — they'll find existing downloads and write new ones back to the same place.
  • Another ComfyUI / single-file loader: point it at ~/ai-models/comfyui via its own extra_model_paths.yaml or a symlink: ln -s ~/ai-models/comfyui /path/to/other/ComfyUI/models.

Relocate the whole store with export AI_MODELS_DIR=/somewhere/else before sourcing env.sh.

Downloading

./download-models.sh list                 # show known targets
./download-models.sh z-image-turbo qwen-2512  # MFLUX: download + tiny test render
./download-models.sh comfy-pony-v7         # ComfyUI: Pony V7 (pixel-art)
./download-models.sh comfy-flux2-dev       # ComfyUI: FLUX.2 dev 32B set
  • MFLUX models auto-download into $HF_HOME on first generate.sh run too.
  • ComfyUI single files are staged and moved into $COMFY_MODELS_DIR (staging is on the same volume, so the move is instant even for the 27 GB GGUF).

See tips-and-gotchas for download reliability (Xet disabled, orphaned .incomplete cleanup, gated-repo licenses).

FLUX.2 dev (ComfyUI) {#flux2-dev-comfyui}

The "big" FLUX.2 [dev] 32B — quality ceiling, best text — runs on 64 GB via ComfyUI-GGUF (MFLUX only supports the klein variants). comfy-flux2-dev fetches:

File Size →
unsloth/FLUX.2-dev-GGUF → flux2-dev-Q6_K.gguf 27.4 GB unet/
Comfy-Org/flux2-dev → Mistral text-encoder fp8 18 GB text_encoders/
Comfy-Org/flux2-dev → flux2-vae.safetensors 0.34 GB vae/
Comfy-Org/flux2-dev → Turbo LoRA 2.76 GB loras/

The unsloth GGUF is ungated. The Turbo LoRA lets the 32B run in ~6–8 steps. Workflow: UnetLoaderGGUF (Q6_K) → CLIP loader (Mistral) → sampler → VAELoader; ComfyUI ships a FLUX.2 template, or use ComfyUI-Manager.

History

models-inventory.md (in this dir) is the archived snapshot of what existed before the models were cleared — kept for reference, superseded by the current models-guide.