Short answer (2026): Installing Flux in ComfyUI is four steps: (1) install the ComfyUI-GGUF custom node if you're low on VRAM, (2) drop the model files into the right folders (diffusion_models/unet, text_encoders, vae), (3) set fp8 in the Load Diffusion Model node (or use a GGUF loader for quantized models), and (4) load a ready-made Flux workflow from the ComfyUI Manager and queue a test.
-
High VRAM (16GB+): fp8
.safetensors+ the standard diffusion loader - Low VRAM (6–8GB): GGUF quant + ComfyUI-GGUF node
- Don't skip: the T5 + CLIP text encoders and the VAE
What you need to download
Flux is not one file — it's four components:
| Component | Folder | Notes |
|---|---|---|
| Diffusion model (Flux.1-dev) |
models/diffusion_models/ (or unet/) |
fp8 .safetensors or a GGUF quant |
| T5 text encoder | models/text_encoders/ |
Use the quantized GGUF T5 on low VRAM (fp16 is ~9GB) |
| CLIP-L encoder | models/text_encoders/ |
clip_l.safetensors |
| VAE | models/vae/ |
ae.safetensors |
For 8GB systems, look for fp8 versions (e.g., huggingface.co/Kijai/flux-fp8) or a GGUF quant.
Path A — fp8 (simplest, for 16GB+ GPUs)
- Put the fp8
flux1-dev.safetensorsinmodels/diffusion_models/. - Add the T5, CLIP-L, and VAE to their folders above.
- In ComfyUI, load the default Flux workflow (Manager → Workflow browser, or drag in a known-good JSON).
- In the Load Diffusion Model node, set
weight_dtype = fp8_e4m3fn. Set fp8 here, in the node — not on the command line. ComfyUI's--fp8_e4m3fn-unetflag is often ignored by Flux's loader. - Queue a prompt.
Path B — GGUF (for 6–8GB GPUs)
- Install ComfyUI-GGUF: Custom Nodes Manager → search "GGUF" → install → restart. (Or
git clone https://github.com/city96/ComfyUI-GGUFintocustom_nodes.) - Download a GGUF model (Q4_K_S is the 8GB sweet spot) into
models/unet/. - Download the quantized GGUF T5 into
models/text_encoders/— not the fp16 one. - Load a GGUF workflow from the Manager's workflow browser; it uses the Unet Loader (GGUF) node instead of the standard loader.
- Launch ComfyUI with
--lowvramand queue a prompt.
Full low-VRAM detail (quant levels, the T5 trap, memory flags) is in our Flux on 8GB VRAM guide. Once it's running, level up your results with the best Flux LoRAs for realism.
Verify it works
Queue a generation with a simple prompt. If an image appears in the preview node with no red error nodes, Flux is installed correctly. Red nodes almost always mean a missing file in one of the four folders above — recheck the T5, CLIP, and VAE first.
Frequently asked questions
Where do Flux model files go in ComfyUI?
The diffusion model goes in models/diffusion_models/ (or unet/ for GGUF), the T5 and CLIP-L encoders in models/text_encoders/, and the VAE in models/vae/.
fp8 or GGUF for Flux — which should I use?
Use fp8 .safetensors if you have 16GB+ VRAM (simplest). Use GGUF with the ComfyUI-GGUF node if you're on 6–8GB — it compresses the model to fit.
Why is my Flux workflow showing red error nodes?
Almost always a missing file. Confirm the diffusion model, T5 encoder, CLIP-L, and VAE are all present in their correct folders — the text encoders and VAE are the most commonly forgotten.
Do I need to set fp8 on the command line?
No — set weight_dtype = fp8_e4m3fn inside the Load Diffusion Model node. The --fp8_e4m3fn-unet CLI flag is frequently ignored by Flux's loader.
Conclusion
Flux in ComfyUI comes down to putting four files in four folders and choosing fp8 (high VRAM) or GGUF (low VRAM). Load a prebuilt workflow from the Manager rather than wiring from scratch, and you'll be generating in minutes. Got a favorite Flux workflow? Share it below.
Top comments (0)