FLUX 3 Image Klein:Generate, Edit & Compose in One Model
FLUX 3 Image Klein is the open-weights member of the FLUX 3 Image family. One 5.6B model handles text-to-image, instruction editing and multi-reference composition — with bounding-box layout control, output up to 4K, and a distilled variant that renders in just 16 steps.
What FLUX 3 Image Klein Can Make
Photography, illustration, posters and infographics — from woodblock, mosaic and clay to pixel art, anime, linocut and cross-stitch. Sample outputs from Black Forest Labs.
Photography
Illustration
Type & Design
One Set of Weights, Every Image Task
Klein draws on the shared learning of the FLUX 3 family across images, video and audio — packed into a model small enough to run on a single consumer GPU.
Trained on full scene captions covering framing, subject, background and lighting. A bundled prompt upsampler turns short prompts into that format automatically.
Describe a change in one line — "put a red bandana around the dog's neck" — and Klein edits that part of the image while keeping the rest intact.
Combine up to 10 reference images into one composition — products, characters, styles and scenes. The first reference sets the output size.
Give each element a box and a description to place it exactly where you want. In edits, boxes can move content from a reference into a target region.
The distilled model runs in 16 steps without guidance — about 6x faster than the base model. On one H200 that's roughly 3 s per 1 MP image versus 12 s.
Klein Base is undistilled for fine-tuning, LoRA training and research. A LoRA trained on Base runs on the distilled model at 16 steps.
Edit One Part, Keep the Rest
Before and after: each edit came from a short text instruction, changing what was asked and preserving everything else.
Lay Out a Scene with Bounding Boxes
Instead of hoping the model places things where you want, give each element a box and a description. The canvas is a normalized grid regardless of aspect ratio, and every box is written as [y_min, x_min, y_max, x_max].
[
{ "id": "moon", "desc": "gold crescent moon",
"bbox": [0.08, 0.30, 0.28, 0.50] },
{ "id": "year", "desc": "gold serif text '1959'",
"bbox": [0.06, 0.62, 0.16, 0.92] },
{ "id": "cat", "desc": "fluffy white cat gazing up",
"bbox": [0.45, 0.10, 0.95, 0.90] }
]Boxes guide placement and scale — they are not hard masks.
Klein vs. Klein Base
Two checkpoints, same 5.6B architecture. Pick distilled for speed, Base for training.
| Klein (Distilled) | Klein Base | |
|---|---|---|
| Parameters | 5.6B | 5.6B |
| Inference steps | 16 | 50 |
| Guidance (CFG) | 1 (none) | 3.5 |
| Max resolution | Up to 4K | Up to 4K |
| Text-to-image / Edit / Boxes | Yes / Yes / Yes | Yes / Yes / Yes |
| Best for | Fast generation & self-hosting | Fine-tuning, LoRA & research |
| Availability | API, Hugging Face | Hugging Face |
| License | FLUX Kommunity License | FLUX Kommunity License |
Local Speed on Real GPUs
| Hardware | 1024² Generate | 1024² Edit | 2048² Generate |
|---|---|---|---|
| RTX 5090 · INT8 distilled | 4.1 s | 7.2 s | 15 s |
| RTX 5090 · INT8 base | 12.5 s | 27.4 s | 79.8 s |
| RTX 4070 · INT8 distilled | 8.9 s | 18.9 s | 50.4 s |
| DGX Station · NVFP4 distilled | 0.59 s | — | 2.2 s |
Figures published by Black Forest Labs; not independently verified.
Runs on consumer GPUs down to an RTX 3090. The INT8 ComfyUI files target 8 GB cards (RTX 3070 class).
FP8 needs Ada, Hopper or Blackwell; NVFP4 needs Blackwell.
A 9B vision-language upsampler rewrites short prompts and edit instructions into full scene captions. BFL says not to run Klein without it.
It adds about 28 GB of GPU memory — or run it once ahead of time on low-VRAM setups.
FLUX Kommunity License: free commercial use for individuals and businesses under US$5M in annual revenue.
Larger organizations need a commercial license from BFL for self-hosting, or can use the API.
API Pricing per Image
The same price for text-to-image and editing on the BFL API.
| Model | Approx. price at 1K | Weights |
|---|---|---|
| FLUX 3 Image Klein | $0.009 | Open |
| FLUX 3 Image | $0.048 | API only |
| Krea 3 | $0.03 | API only |
| Nano Banana 2.1 | $0.0336 (output) | API only |
Launch list prices as reported in October 2026. Check each provider for current rates.
Who Klein Is For
Product shots, lifestyle scenes and campaign variations — then edit details without regenerating.
Explore styles from linocut to anime, and train a LoRA on Klein Base to lock in your own look.
Posters, infographics and book covers, with bounding boxes to control where each element lands.
Self-host open weights, run it in ComfyUI, or call the API from $0.008 per image.
Frequently Asked Questions
Common questions about FLUX 3 Image Klein and how to get started.
About the numbers. Specs, pricing and speed figures on this page come from Black Forest Labs' launch materials and press coverage. Quality and speed claims are the vendor's own and have not been independently verified.
Ready to Create withFlux Image Models?
Skip the 100 GB download and the CUDA setup. Generate and edit Flux images in your browser in seconds.