On this page
FLUX.2 [dev] is Black Forest Labs' 32B open-weight image model. It generates and edits from several reference images at once. Here's the shortest path to running it in ComfyUI on one card.
- Generator
- 32B · text-to-image + multi-reference edit
- Text encoder
- Mistral Small 3 · 24B
- Output
- up to 4 MP
- Licence
- FLUX [dev] non-commercial
Pick your GPU
Choose once — every block below follows your pick.
NVFP4 — BFL's 4-bit build, made for Blackwell. The generator and encoder don't fit together, so ComfyUI loads them in turn.
| File | Size |
|---|---|
| Generator · NVFP4 mixed | 22.8 GB |
| Text encoder · FP8 | 18.0 GB |
The template's default flux2_dev_fp8mixed (35.5 GB) also runs — ComfyUI offloads what doesn't fit, so it's slower.
Install ComfyUI
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
python3 -m venv .venv && source .venv/bin/activate
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt "huggingface_hub[cli]"Download the models
Run from the ComfyUI folder. Every file lands in the right models/ folder.
hf download black-forest-labs/FLUX.2-dev-NVFP4 flux2-dev-nvfp4-mixed.safetensors --local-dir models/diffusion_models
hf download Comfy-Org/flux2-dev split_files/text_encoders/mistral_3_small_flux2_fp8.safetensors --local-dir models/.hf
hf download black-forest-labs/FLUX.2-small-decoder full_encoder_small_decoder.safetensors --local-dir models/vae
hf download ByteZSzn/Flux.2-Turbo-ComfyUI Flux_2-Turbo-LoRA_comfyui.safetensors --local-dir models/loras
mv models/.hf/split_files/text_encoders/* models/text_encoders/Launch
python main.pyOpen http://127.0.0.1:8188.
Generate
Workflow → Browse Templates → Image → Flux.2 Dev (with references) or Flux.2 Dev Text to Image. Then point three loaders at your files:
- Load Diffusion Model —
flux2-dev-nvfp4-mixedon the 5090,flux2-devon the PRO 6000. - Load CLIP —
mistral_3_small_flux2_fp8. - Load VAE — already
full_encoder_small_decoder.
Defaults are 20 steps at guidance 4. Flip the template's turbo switch to true and it drops to 8 steps using the turbo LoRA.
References: each Load Image feeds a ReferenceLatent node. Chain another pair for every extra image, then say in the prompt what each one is for:
The woman from the first image, wearing the jacket from the second image.
Standing on a rain-wet street at night, neon signs reflected in puddles.
85 mm portrait, shallow depth of field, soft rim light.My results
Same prompt and seed on both cards — filling this in as I test.
| Card | Build | 1 MP · 20 steps | 1 MP · turbo 8 | Peak VRAM |
|---|---|---|---|---|
| RTX 5090 | NVFP4 mixed | — | — | — |
| RTX 5090 | FP8 mixed | — | — | — |
| RTX PRO 6000 | Full BF16 | — | — | — |