FLUX 3 Klein in ComfyUI: the Minimum Setup
What you need
- GPU — an NVIDIA RTX card with 8GB VRAM or more. BFL states Klein runs on as little as 8GB through ComfyUI using the quantized build; 16GB+ lets you run the full-precision weights.
- ComfyUI — updated to a build with FLUX 3 support. ComfyUI shipped day-one support, so an update of your existing install is usually all it takes.
- The weights — from the ComfyUI-optimized repo on Hugging Face (flux-3-image-klein-comfy). Two sizes exist, and picking the right one is the whole setup.
Which weight file for which GPU
| Your VRAM | Download | Why |
|---|---|---|
| 8–12 GB | FP8 build (~5.6 GB) | fits the card with headroom for the text encoder and a 1MP latent |
| 16 GB+ | BF16 build (~11.2 GB) | full precision; marginal quality gain, no quantization artifacts |
| Under 8 GB | — | not a local target; use a free cloud option instead |
Sizes are from BFL's announcement (11.2 GB BF16 / 5.6 GB FP8). Download the one matching your card — downloading both wastes 17 GB for nothing.
The minimum workflow
ComfyUI's official docs ship a FLUX 3 Klein template (docs.comfy.org) — load that instead of wiring nodes by hand. The shape is the standard one: load the model → encode your prompt → sample → decode. Two Klein-specific points:
- Klein is the distilled variant, so it needs fewer sampling steps than the undistilled Klein Base. Start from the template defaults; don't copy step counts from FLUX.1 workflows.
- Multi-reference editing is Klein's headline feature — it accepts a prompt plus up to 10 reference images. The text-to-image template won't expose that; use the editing template for it.
What speed to expect
BFL's figure: about 4 seconds for a 1MP image on NVIDIA GeForce RTX hardware. That's the reference point for 'is my setup healthy' — if a 1024×1024 generation takes a minute on a modern card, something is wrong (usually VRAM spilling to system RAM, see below).
The three errors everyone hits
- Out of memory / generation crawls — you loaded the BF16 build on an 8–12 GB card, or your resolution is too high. Switch to FP8, drop to 1MP, close other GPU apps.
- Missing or red nodes — ComfyUI is older than FLUX 3 support. Update ComfyUI (and the manager) before anything else.
- Model fails to load — an incomplete download. Compare the file size against the Hugging Face listing and re-download; partial safetensors files fail silently until load time.
After it works
Grab the prompt pack to see what Klein is good at, and read the Klein vs Klein Base comparison before you consider fine-tuning — Base, not Klein, is the one meant for that.