---
language:
- en
license: other
library_name: transformers
---

**Flux assets for ComfyUI**

This repository contains **reference weights** arranged for ComfyUI under:

```
models/
  diffusion_models/
  vae/
  text_encoders/
```

When used with Hugging Face **Inference Endpoints**, these files are baked into the endpoint image and available at:

```
/repository/models/...
```

Configure ComfyUI (`extra_model_paths.yaml`) to include the paths above.

> **Important:** Check the **original model licenses and terms** (Black Forest Labs / Comfy-Org / comfyanonymous repos). Some weights **do not permit commercial or production use**. This repository is provided strictly for **demo/testing** purposes; you are responsible for ensuring compliance before any other use.

**Contents:**

* `models/diffusion_models/` — Flux checkpoint (e.g., `flux1-krea-dev_fp8_scaled.safetensors`)
* `models/vae/` — VAE (`ae.safetensors`)
* `models/text_encoders/` — CLIP/T5 encoders

If you need a different layout, adjust `extra_model_paths.yaml` accordingly.


**Hugging Face Inference Endpoints**

- set the health route to `/`.
- set COMFY_FLAGS

**Docker (explicit flags):**

```bash
docker run --rm --gpus all -p 8080:80 \
  -e COMFY_FLAGS="--normalvram --use-pytorch-cross-attention --cache-lru 64 --reserve-vram 1.5" \
  your-image:tag
```

**Recommended `COMFY_FLAGS`**

These presets prioritize stability. They’re meant as safe defaults; you should still **set `COMFY_FLAGS` explicitly** for your deployment when you know your workload.

`--disable-xformers --use-pytorch-cross-attention --cache-lru 2 --reserve-vram 2.0`

**Minimal recommended `hardware`**

- Nvidia L4: 8 vCPU, 32 GiB RAM, disk 40 GiB