How to Install Flux AI Locally in ComfyUI
Flux, developed by Black Forest Labs, is one of the most capable open-weight text-to-image model families available. This guide covers the full lineup as of mid-2026: the original Flux.1 Schnell and Dev, and the newer Flux.2 generation (Klein and Dev), along with which VRAM tier and quantization format fits your GPU.
Before you start, you'll need:
- ComfyUI installed and updated to the latest version — see our ComfyUI install guide if you haven't yet.
- ComfyUI Manager installed — makes downloading missing nodes and updating far easier.
- An NVIDIA GPU with at least 8 GB VRAM (see the VRAM guide below for exactly which variant fits your card).
- Roughly 15–65 GB of free disk space per model, depending on which variant and precision you choose.
Why Flux Replaced SDXL as the Go-To Model
SDXL was the default choice for local image generation through 2024. Flux overtook it because it follows prompts far more literally — including counting, spatial relationships ("to the left of"), and multi-subject scenes — and because it renders readable text inside images, something SDXL consistently failed at.
SDXL still wins on one thing: the LoRA and fine-tune ecosystem built up over years is larger and more mature. If your workflow depends on a specific character or style LoRA that only exists for SDXL, that's a real reason to stay. For everything else — photorealism, prompt accuracy, and text — Flux is the stronger default in 2026, and Flux 2 widens that gap further with better world-physics reasoning and multi-reference support.
Flux Model Variants Explained
Flux comes in two generations. Flux.1 (Schnell, Dev, Pro) is the original lineup. Flux.2 (Klein, Dev) is the newer generation with better detail, multi-reference support, and — with Klein — a genuinely fast, commercial-friendly small model. Here's how all of them compare:
| Model | License | VRAM (best case) | Speed | Quality | Price |
|---|---|---|---|---|---|
| Flux.1 Schnell | Apache 2.0 | 16 GB | ⚡ Very Fast | ★★★☆ | Free |
| Flux.1 Dev | Non-commercial | 24 GB (12 GB GGUF Q8) | 🐢 Moderate | ★★★★ | Free |
| Flux.2 Klein 4B | Apache 2.0 | 12 GB | ⚡ Very Fast | ★★★★ | Free |
| Flux.2 Klein 9B | Apache 2.0 | 16 GB | ⚡ Fast | ★★★★☆ | Free |
| Flux.2 Dev | Non-commercial | 32 GB (19 GB GGUF Q4) | 🐢 Slow | ★★★★★ | Free |
| Flux.1 Pro | API Only | Cloud | ⚡ Fast | ★★★★ | $0.05/img |
| Flux1.1 Pro | API Only | Cloud | ⚡ Fast | ★★★★★ | $0.04/img |
Free Models (Local Install)
- Flux.1 Schnell: Fast, Apache 2.0 licensed — free for personal and commercial use. Needs ~16 GB VRAM.
- Flux.1 Dev: Higher quality, non-commercial license. Needs ~24 GB VRAM (or ~12 GB as a GGUF Q8).
- Flux.2 Klein 4B: Apache 2.0, distilled for speed. Runs on 12 GB cards, sub-second generation.
- Flux.2 Klein 9B: Apache 2.0, sharper detail than 4B. Fits 16 GB cards at fp8.
- Flux.2 Dev: The largest, highest-quality Flux model (32B params), non-commercial. Needs ~32 GB at fp8 or ~19 GB as a GGUF Q4 on a 24 GB card.
Paid Models (API Only)
- Flux.1 Pro: High-speed, high-quality via API — $0.05 per image.
- Flux1.1 Pro: Best overall API performance — $0.04 per image.
VRAM & Quantization Guide: Full, fp8, GGUF, and NF4
Every Flux model above is listed at its "best case" VRAM — meaning a quantized version. Quantization shrinks the model file so it fits smaller GPUs, at a small, usually hard-to-notice quality cost. Here's what each format actually means:
| Format | Typical VRAM | Quality vs. full | Node needed | Best for |
|---|---|---|---|---|
| BF16 / FP16 (full) | 24–64 GB | Baseline (100%) | Load Diffusion Model | Cards with 24 GB+ and no compromise on quality |
| FP8 (e4m3fn) | 12–32 GB | Very close to full | Load Diffusion Model (set weight_dtype) | 16 GB+ cards — the default choice for most people |
| GGUF (Q8) | ~50% of full | Near-identical to FP8 | Unet Loader (GGUF) | 12 GB cards, or anyone who wants fine control over size |
| GGUF (Q4–Q5) | ~25–35% of full | Slightly softer on complex prompts | Unet Loader (GGUF) | 6–8 GB cards — the only realistic path at this tier |
| NF4 | ~25% of full | Comparable to GGUF Q4 | Community NF4 loader node | Forge/A1111 users; less standardized in ComfyUI than GGUF |
fp8 — the default for 16 GB+ cards
fp8 halves the file size of the full BF16/FP16 model with minimal quality loss. It uses the same Load Diffusion Model node as the full-precision model — just set weight_dtype to fp8_e4m3fn inside the node. This is the simplest option if your card has enough VRAM, and it's what most of the download links in this guide point to by default.
GGUF — the standard low-VRAM path
GGUF compresses the model further, with selectable compression levels from Q8 (near-fp8 quality) down to Q4 (smallest, for 6–8 GB cards). It requires the community ComfyUI-GGUF custom node, which adds a new Unet Loader (GGUF) node — this replaces Load Diffusion Model in your workflow. Install it from ComfyUI Manager by searching "GGUF," or via git clone https://github.com/city96/ComfyUI-GGUF custom_nodes/ComfyUI-GGUF. You'll also need the matching GGUF version of your text encoder — not the fp16 one.
NF4 — a less common alternative
NF4 (4-bit NormalFloat) is another 4-bit format, more associated with Forge and A1111 through the bitsandbytes library than with ComfyUI. Community NF4 loader nodes exist for ComfyUI, but they're less mature and less widely tested than GGUF. Unless you already have an NF4 workflow you're attached to, GGUF is the safer, better-supported choice for low-VRAM Flux in ComfyUI today.
Installing Flux.1 Schnell on ComfyUI
Flux.1 Schnell is the fastest option for local image generation. It produces good quality results in fewer steps and is available under the Apache 2.0 license. You will need 16 GB of VRAM to run it effectively at full precision.
Image Generated Using Flux.1 Schnell
Sample Prompt
A young artistic Japanese woman with tousled bangs, smudged eyeliner, and a focused expression of creative absorption, dressed in a patchwork denim jacket over a sheer mesh top, high-waisted cargo skirt, and rugged combat boots, snaps a playful mirror selfie inside a vibrant purikura booth, her cheeks puffed out dramatically while surrounded by floating cartoon sparkles, digital hearts, and kawaii stickers on the glowing screen. The booth opens to a bustling retro arcade hallway where rainbow neon lights flash in rapid sequence, reflecting vividly in the wide eyes of excited passersby and glossy storefront displays filled with anime merchandise. Photorealistic cyberpunk style, wide-angle shot from inside the booth using a 35mm prime lens, dynamic neon lighting with high contrast and colorful glows casting dramatic shadows, ultra-detailed textures on fabrics, skin, and digital effects, sharp focus on the woman's face and screen with soft bokeh in the arcade background.
Installing Flux.1 Dev on ComfyUI
Flux.1 Dev produces higher-quality images than Schnell and follows prompts more precisely. It is licensed for non-commercial use and needs 24 GB of VRAM at full precision — or around 12 GB using a GGUF Q8 quant (see the low-VRAM section below).
Image Generated Using Flux.1 Dev
Sample Prompt
A young artistic Japanese woman with tousled bangs, smudged eyeliner, and a focused expression of creative absorption, dressed in a patchwork denim jacket over a sheer mesh top, high-waisted cargo skirt, and rugged combat boots, snaps a playful mirror selfie inside a vibrant purikura booth, her cheeks puffed out dramatically while surrounded by floating cartoon sparkles, digital hearts, and kawaii stickers on the glowing screen. The booth opens to a bustling retro arcade hallway where rainbow neon lights flash in rapid sequence, reflecting vividly in the wide eyes of excited passersby and glossy storefront displays filled with anime merchandise. Photorealistic cyberpunk style, wide-angle shot from inside the booth using a 35mm prime lens, dynamic neon lighting with high contrast and colorful glows casting dramatic shadows, ultra-detailed textures on fabrics, skin, and digital effects, sharp focus on the woman's face and screen with soft bokeh in the arcade background.
Installing Flux.2 Klein on ComfyUI (New for 2026)
Flux.2 Klein is the newest addition to the lineup — a distilled model built for speed, released in two sizes: 4B (for 12 GB cards) and 9B (for 16 GB cards). It's Apache 2.0 licensed, so unlike Flux.1 Dev, you can use the output commercially. If you want the best free Flux experience on a mid-range GPU in 2026, this is it.
Image Generated Using Flux.2 Klein
Sample Prompt
A young artistic Japanese woman with tousled bangs, smudged eyeliner, and a focused expression of creative absorption, dressed in a patchwork denim jacket over a sheer mesh top, high-waisted cargo skirt, and rugged combat boots, snaps a playful mirror selfie inside a vibrant purikura booth, her cheeks puffed out dramatically while surrounded by floating cartoon sparkles, digital hearts, and kawaii stickers on the glowing screen. The booth opens to a bustling retro arcade hallway where rainbow neon lights flash in rapid sequence, reflecting vividly in the wide eyes of excited passersby and glossy storefront displays filled with anime merchandise. Photorealistic cyberpunk style, wide-angle shot from inside the booth using a 35mm prime lens, dynamic neon lighting with high contrast and colorful glows casting dramatic shadows, ultra-detailed textures on fabrics, skin, and digital effects, sharp focus on the woman's face and screen with soft bokeh in the arcade background.
Running Flux on Low VRAM (6–8 GB) with GGUF
If your card doesn't meet the VRAM numbers above, GGUF is the way to run Flux.1 Dev or Flux.2 Klein anyway. This applies to any Flux variant — swap in the GGUF build of whichever model you chose above.
Troubleshooting
Red nodes when you load the workflow
Red nodes mean missing custom nodes. Open Manager → Install Missing Custom Nodes → restart ComfyUI. If they're still red, make sure ComfyUI itself is updated first (Manager → Update All) — outdated cores are the most common cause.
"CUDA out of memory" during generation
Your VRAM tier is too small for the model you loaded. Switch to a GGUF build (see the low-VRAM section above), lower your output resolution, or launch ComfyUI with --lowvram. Also check nothing else is holding VRAM — close browsers with hardware acceleration and other GPU apps.
Blurry, washed-out, or heavily distorted output
Almost always a VAE mismatch, or CFG set too high. Confirm you're using the correct VAE for your model family (Flux.1 vs. Flux.2 use different VAEs) and that CFG is set to 1.0, not 5–7 like a Stable Diffusion workflow.
Model doesn't appear in the loader dropdown after downloading
Confirm the file is in the exact folder the node expects — diffusion_models for the main model, text_encoders for CLIP/T5/Qwen, vae for the VAE. Then refresh ComfyUI in the browser (press R) or fully restart it — new files aren't picked up live.
Frequently Asked Questions
Ready to go further? Check the ComfyUI roadmap for the recommended next steps after installing Flux.
Join the discussion
Sign in to leave a comment or reply
No comments yet
Be the first to share your thoughts!