Wonderful Launcher 文档ComfyUI Flux Guide: Setup, Workflows & VRAM Options
Complete guide to running Flux.1 in ComfyUI — model versions compared, download links, text-to-image and image-to-image workflows, and low-VRAM solutions.
What is Flux.1?
Flux.1 is a text-to-image model family created by Black Forest Labs (founded by Robin Rombach, former core member of Stability AI). It delivers exceptional prompt following, text rendering inside images, and natural hand anatomy — areas where Stable Diffusion models traditionally struggle.
Model Versions
| Version | Quality | License | Download |
|---|---|---|---|
| Flux.1 Pro | Top-tier | API only, commercial | Not downloadable |
| Flux.1 Dev | Near-Pro | Non-commercial, open-weight | HuggingFace |
| Flux.1 Schnell | Good | Apache 2.0, commercial OK | HuggingFace |
Hardware Requirements
| Setup | VRAM | Notes |
|---|---|---|
| Original (bf16/fp16) | 16–24 GB | Best quality, requires powerful GPU |
| FP8 Checkpoint | 8–12 GB | Good balance of quality and performance |
| GGUF Quantized | 6–8 GB | Accessible to most consumer GPUs |
| NF4 | 6 GB+ | Smallest footprint, requires plugin |
Choose Your Setup Path
Flux has multiple packaging formats. Pick the one that matches your hardware:
Path A: FP8 Checkpoint (Recommended for Most Users)
The simplest setup — one file, no extra plugins needed.
Download one model:
| Model | Size | Download |
|---|---|---|
| flux1-dev-fp8 | 17.2 GB | HuggingFace |
| flux1-schnell-fp8 | 17.2 GB | HuggingFace |
Place in ComfyUI/models/checkpoints/.
Workflow: Use a standard Load Checkpoint → CLIP Text Encode → KSampler workflow, just like SD1.5.
Path B: Original Multi-File Setup
Best quality, but requires downloading multiple files:
CLIP models (place in ComfyUI/models/clip/):
| File | Size | Download |
|---|---|---|
| clip_l.safetensors | 246 MB | HuggingFace |
| t5xxl_fp8_e4m3fn.safetensors | 4.89 GB | HuggingFace |
VAE (place in ComfyUI/models/vae/):
| File | Size | Download |
|---|---|---|
| ae.safetensors | 335 MB | HuggingFace |
UNET (place in ComfyUI/models/unet/):
| File | Size | Download |
|---|---|---|
| flux1-dev.safetensors | 23.8 GB | HuggingFace |
| flux1-schnell.safetensors | 23.8 GB | HuggingFace |
This workflow uses separate loader nodes: UNET Loader, Dual CLIP Loader, and VAE Loader instead of a single Load Checkpoint.
Path C: GGUF (Lowest VRAM)
Requires the ComfyUI-GGUF plugin.
| File | Where | Download |
|---|---|---|
| GGUF model (choose Q4/Q5/Q8 by quality) | ComfyUI/models/unet/ | Dev GGUF / Schnell GGUF |
| GGUF text encoder | ComfyUI/models/clip/ | t5-v1_1-xxl-encoder-gguf |
| clip_l.safetensors | ComfyUI/models/clip/ | Download |
| ae.safetensors | ComfyUI/models/vae/ | Download |
GGUF files come in different quantization levels: Q2 (smallest, lowest quality) through Q8 (largest, highest quality). Start with Q4_K_M for a good balance.
Text-to-Image Workflow
For the FP8 Checkpoint path, the workflow is straightforward:
- Load Checkpoint → select your flux1-dev-fp8 or flux1-schnell-fp8 model
- CLIP Text Encode → write your prompt (Flux has excellent text understanding — write naturally)
- Empty Latent Image → set resolution (recommended: 1024x1024 or similar 1MP resolutions)
- KSampler → steps: 20 (Dev) or 4 (Schnell), cfg: 1.0, sampler: euler
- VAE Decode → Save Image
Flux uses a cfg of 1.0 — this is different from SD1.5/SDXL. Higher cfg values will produce over-saturated, distorted images.
Image-to-Image Workflow
Replace the Empty Latent Image with:
- Load Image → load your source image
- VAE Encode → converts the image to latent space
Connect the VAE Encode output to the KSampler's latent input. Adjust denoise (0.3–0.7) to control how much the image changes.
Flux Redux: Image Variation Without Prompts
Flux Redux generates style variants from a reference image — no text prompt required. It needs additional model files:
| File | Location | Download |
|---|---|---|
| sigclip_vision_patch14_384.safetensors | models/clip_vision/ | Download |
| flux1-redux-dev.safetensors | models/style_models/ | Download |
Flux Ecosystem
Flux has a growing ecosystem of add-ons:
| Type | Resources |
|---|---|
| LoRA | Civitai Flux LoRAs, XLabs collection |
| ControlNet | XLabs ControlNet, InstantX Union |
| IP-Adapter | XLabs IP-Adapter |
| LoRA Training | FluxGym (local, low VRAM) |
Windows Optimization
If you're running Flux on Windows with limited RAM:
- Increase Windows virtual memory (page file) to at least 32 GB
- Close other applications while generating
- Use the FP8 or GGUF versions to reduce memory pressure
Common Issues and Fixes
Out of memory
- Switch to the FP8 Checkpoint version (8 GB VRAM minimum)
- Or use GGUF Q4 (6 GB VRAM minimum)
- Reduce resolution below 1024x1024
Colors are over-saturated or image looks distorted
- Set
cfgto 1.0 — Flux does not use high cfg values like SD1.5
Slow generation speed
- Schnell is 5–10x faster than Dev (4 steps vs 20 steps)
- Use FP8 quantization for faster inference
- See the ComfyUI performance guide
LoRA causing VRAM issues
- Multiple Flux LoRAs can spike VRAM usage significantly
- Consider using the GGUF model when combining LoRAs
Related Guides
- Text to Image — Basic image generation concepts
- LoRA Guide — How to use LoRAs in ComfyUI
- ControlNet Guide — Structural control over generation
ComfyUI Model Types Explained: Checkpoint, LoRA, VAE, ControlNet & More
Understand the different model types used in ComfyUI — what each does, where to install it, and how they work together in a workflow.
ComfyUI Flux Kontext Guide: AI Image Editing with Text Instructions
How to use FLUX.1 Kontext in ComfyUI for context-aware image editing — change objects, transfer styles, edit text in images, and maintain character consistency.