Black Forest Labs has released FLUX 3 Image Klein, the compact open-weights member of the FLUX 3 family, and it runs natively in ComfyUI from the day of release.
At 5.6 billion parameters it is the model in the family intended to be iterated on by one person with one GPU, and BFL says it draws on learning shared across FLUX 3’s image, video and audio models while keeping the control features of the much larger FLUX 3 Image — bounding boxes included.
Two checkpoints, and the distinction matters
It ships in two forms, and picking the wrong one wastes a weekend:
- Base — the full undistilled checkpoint, for fine-tuning, further training and research
- Distilled — a lower-latency “turbo” variant for fast generation, suffixed
-sd
If you intend to train a LoRA or a style adapter, you want base. Distilled checkpoints have had their sampling trajectory compressed, which is what makes them fast and also what makes them awkward to fine-tune. If you just want images out, distilled is the one.
The speed numbers are the story
BFL’s published performance test pairs INT8 ConvRot quantisation with the step-distilled checkpoint and reports 4.3× to 9.3× faster 1024×1024 text-to-image generation than the BF16, 50-step baseline.
The gains skew towards cheaper hardware, which is unusual and worth noticing:
| GPU | Before | After |
|---|---|---|
| RTX 5070 | 71 seconds | under 8 seconds |
| RTX 4070 | 78 seconds | under 9 seconds |
A 71-second generation is a model you consult. An 8-second generation is a model you work with — you can hold an idea in your head across the wait. That threshold, not the parameter count, is what decides whether a local model becomes part of a practice.
The reason mid-range cards benefit most is that they are the ones where the BF16 model was memory-bound rather than compute-bound; quantising the weights relieves the bottleneck that was actually binding.
What it can do
- Native 4K — generated at 4K, not upscaled into it
- Prompt upsampler — a separate model that expands a short prompt into a detailed one before generation
- Multi-reference generation and editing — feed it several references; characters and subjects stay recognisable across edits and across inputs
- Bounding boxes — mark regions on the canvas to set composition and layout, deciding where things go before the model fills them in
- Text rendering and editing — writes text into the image, and edits text already there
- Reference-based style transfer — takes the look of a reference while keeping your content and composition
Bounding boxes are the feature to try first if you have been fighting prompts. Composition is the thing text is worst at specifying — “on the left” is a weak instruction and a drawn rectangle is a strong one — and inheriting that control from the full FLUX 3 Image model is the most practically useful thing about this release.
On the benchmark claim
BFL says that in its internal evaluations, FLUX 3 Image Klein scores above Qwen-Image 2.1 on text-to-image and multi-reference generation.
Read that as the vendor’s own number, because it is. Internal evaluations with undisclosed prompts and undisclosed judging are a legitimate thing for a lab to publish and not a thing to plan around. Qwen-Image 2.1 is a reasonable comparison point — it is the model whose quantised repackagings have been climbing Hugging Face all month — but the only evaluation that settles it for your work is your own prompts on your own hardware.
The first comment under ComfyUI’s own announcement, from a user who tried both checkpoints, was that eyes came out distorted. One comment is not a finding either. It is a reminder that day-one impressions of an image model are mostly a function of which prompts someone happened to run.
Getting it running
- Update ComfyUI to the latest version, or use Comfy Cloud
- Download the weights from Hugging Face and pick base or distilled (
-sd), dropping them in your models folder - Load the FLUX 3 Image Klein template from the Templates panel
- Write a prompt, attach reference images if editing, and run
The template is the part that saves time. Multi-reference editing and bounding-box conditioning both need specific node wiring, and starting from a working graph rather than assembling one is the difference between an afternoon and ten minutes.
Sources
- FLUX 3 Image Klein in ComfyUI: 5.6B Open Weights — ComfyUI Blog
- black-forest-labs/flux-3-image-klein-comfy — weights on Hugging Face
- Black Forest Labs
- Black Forest Labs’ open-weights image model FLUX 3 Image Klein runs on consumer GPUs with just 8 GB of VRAM — The Decoder
- Inside Flux 3 Image Klein with BFL! — ComfyUI on YouTube