black-forest-labs/flux-3-image-klein

FLUX 3 Image Klein

Model page • Weights • Demo

This repository contains the reference inference code for FLUX 3 Image Klein, the 5.6B open weights image model from Black Forest Labs: text-to-image, editing with up to 10 reference images, and layout control with bounding boxes, through one model. It ships with a prompt upsampler; do not use the model without it.

Performance

Leads in quality and speed among open weights models: the Pareto frontier of Elo rating against latency, for text-to-image and for image editing.

_{Elo from our internal evaluations; median API latency for a 1 MP image.}

Model overview

| Name | Steps | Guidance | Resolution | Text-to-image | Editing (up to 10 references) | Bounding boxes | License | | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | FLUX 3 Image Klein | 16 | none | up to 4K | ✅ | ✅ | ✅ | FLUX Kommunity License | | FLUX 3 Image Klein Base | 50 | CFG 4 | up to 4K | ✅ | ✅ | ✅ | FLUX Kommunity License | | FLUX 3 Image Klein Prompt Upsampler | | | | | | | Apache 2.0 |

Which model should I use?

| Need | Recommended | |------|-------------| | Generation and editing, self-hosting | FLUX 3 Image Klein (distilled): the same model in 16 steps without CFG, about 6x faster | | Fine-tuning, LoRA training, research | FLUX 3 Image Klein Base | | An 8 GB GPU (RTX 3070 class) | the int8 files in ComfyUI |

[!IMPORTANT]

Do not use either model without the prompt upsampler. The model is trained on the captions it

writes; a prompt sent as typed gives noticeably weaker results. It is on by default in the CLI, the demo and the

Python API.

Text-to-image examples

Editing examples

Layout examples

Typography

Base vs distilled

Same prompt, same seed: distilled (16 steps) left, base (50 steps, CFG) right. On one H200 at 1 MP: about 3 s against 12 s per image.

Local installation

To run FLUX 3 Image Klein in BF16 and FP8:

python3.12 -m venv .venv && . .venv/bin/activate
pip install -e . --extra-index-url https://download.pytorch.org/whl/cu130

To run FLUX 3 Image Klein in BF16, FP8 and NVFP4 (NVFP4 needs a Blackwell GPU):

python3.12 -m venv .venv && . .venv/bin/activate
pip install -e '.[nvfp4]' --extra-index-url https://download.pytorch.org/whl/cu130 --extra-index-url https://flashinfer.ai/whl/cu130

The installation was tested on B200; see Memory for other GPUs.

Weights

The fast model with its prompt upsampler, 52 GB:

hf download black-forest-labs/flux-3-image-klein --local-dir weights/flux-3-image-klein --exclude 'variants/*'
hf download black-forest-labs/flux-3-image-klein-pup --local-dir weights/pup
hf download Falconsai/nsfw_image_detection --local-dir weights/nsfw-classifier

Everything, 107 GB:

hf download black-forest-labs/flux-3-image-klein-base --local-dir weights/flux-3-image-klein-base
hf download black-forest-labs/flux-3-image-klein --local-dir weights/flux-3-image-klein
hf download black-forest-labs/flux-3-image-klein-pup --local-dir weights/pup
hf download Falconsai/nsfw_image_detection --local-dir weights/nsfw-classifier

WEIGHTS_DIR points at another root (default ./weights). The quantized DiTs live under variants/fp8r and variants/nvfp4 in each repository; the text encoder and autoencoder ship with both. weights/pup is the prompt upsampler, which the model needs.

Run the CLI

[!IMPORTANT]

The commands below send the prompt through the upsampler before the model. Do not pass

--no-pup unless the prompt already is an upsampler caption (Smaller GPUs).

Text to image

python scripts/cli.py "A man walking with his dog, afternoon sun" --model distilled --out out.png --seed 7

Each generated image comes with a .json containing the prompt, size, number of steps, guidance value and seed; examples/README.md has the command behind every image in this README.

Edit an image

python scripts/cli.py "put a red bandana around the dog's neck" --image examples/inputs/3.png --out out.png --seed 22

Up to 10 references; the first one sets the output size:

python scripts/cli.py "print the dog from the second image onto the mug" \
  --image examples/inputs/2.png --image examples/inputs/3.png --out out.png --seed 23

Place elements with bounding boxes

python scripts/cli.py "hot-air balloons over a green valley at sunrise" \
  --boxes examples/inputs/balloons_boxes.json --out out.png --seed 1

Format in Bounding boxes.

Smaller GPUs

--fp8 (Ada, Hopper, Blackwell) or --nvfp4 (Blackwell) load the quantized DiTs.

An external vision-language model can also write the caption in place of the upsampler: give it the system prompts and pass its answer with --no-pup.

When the upsampler and the model do not fit on the GPU together, run the upsampler on its own first and pass its caption with --no-pup:

python -m flux3.pup --prompt "A man walking with his dog, afternoon sun" --out prompt.txt
python scripts/cli.py "$(cat prompt.txt)" --model distilled --fp8 --no-pup --out out.png

Numbers in Memory.

From Python

from PIL import Image
from flux3.generate import Models, generate
from flux3.pipeline import WEIGHTS, dit_path
models = Models({"distilled": str(dit_path("distilled"))}, pup=str(WEIGHTS / "pup"), pup_memory=0.35)
image, record = generate(models, "distilled", "a red bicycle leaning on a white wall", images=[], seed=1)
image.save("bicycle.png")
image, record = generate(models, "distilled", "put a red bandana around the dog's neck",
                         images=[Image.open("examples/inputs/3.png")], seed=22)
image.save("bandana.png")

Models loads once; generate also takes boxes, width, height, steps, guidance, pup, safety.

Demo

python scripts/app.py            # http://127.0.0.1:7860; --share for a public link

Three tabs: Layout (draw boxes, type what goes in each), Text to image, Edit. --offload, --fp8, --nvfp4 and --no-pup apply as in the CLI.

Prompt upsampling

The upsampler (black-forest-labs/flux-3-image-klein-pup, a 9B vision-language model fine-tuned from Qwen3.5-9B) expands a short request into the detailed prompt the model works best with, before it reaches the text encoder. It is on by default in the CLI, the demo and generate. Do not use the model without it, for text-to-image, editing and layout alike: --no-pup sends the prompt as typed, and the results are noticeably weaker.

| input prompt | what the model saw | |---|---| | double exposure portrait: a young East Asian man facing the camera, a larger translucent profile of him facing left layered over him, muted blues, film grain, silver hoop earring | "Digital double-exposure portrait in a muted blue monochrome palette, framed as a head-and-shoulders composition with visible 35mm film grain. Centered in the foreground, a young East Asian man in his twenties faces the camera directly ... a polished silver hoop earring hangs from his left ear. His face remains sharply defined beneath a larger translucent profile of the same man layered across it, facing left ..." (214 words, examples/upsampling/double_exposure_s1.json) |

python scripts/cli.py "double exposure portrait: a young East Asian man facing the camera, a larger translucent profile of him facing left layered over him, muted blues, film grain, silver hoop earring" \
  --model distilled --width 832 --height 1248 --seed 1 --out out.png

Upsample without generating:

python -m flux3.pup --prompt "double exposure portrait: a young East Asian man facing the camera, a larger translucent profile of him facing left layered over him, muted blues, film grain, silver hoop earring" --out prompt.txt

Bounding boxes

boxes.json is a list of boxes, [y_min, x_min, y_max, x_max] in 0 to 1, top-left origin:

[{"id": "obj_1", "desc": "a red hot-air balloon", "bbox": [0.08, 0.10, 0.45, 0.42]},
 {"id": "obj_2", "desc": "a lighthouse on a cliff", "bbox": [0.30, 0.60, 0.95, 0.90]}]

For edits, a box is {"id", "desc", "from": "ref_image_0" | "target", "src_bbox", "tgt_bbox"}: where it is in a reference, where it goes in the output, or both.

Run it elsewhere

| | | |---|---| | ComfyUI | black-forest-labs/flux-3-image-klein-comfy: bf16, int8 and NVFP4 files for both models. The int8 files run on 8 GB GPUs such as the RTX 3070 | | Diffusers | FLUX 3 Image Klein pipeline docs: modular pipeline, runs the prompt upsampler on every request | | SGLang | FLUX.3 cookbook | | vLLM | vLLM recipe | | TensorRT LLM | Deployment guide: bf16, FP8 and NVFP4 for both models, on the feat/flux3-image-klein branch |

Benchmarks

Seconds per image, one image per run, fastest build per GPU: NVFP4 on Blackwell data-center and DGX parts, int8 on the RTX cards (the ComfyUI files, Run it elsewhere).

FLUX 3 Image Klein, 16 steps

FLUX 3 Image Klein Base, 50 steps

Memory

Peak GPU memory (nvidia-smi) on a B200 at 1024x1024, text-to-image and an edit with one reference:

| | bf16 | --fp8 | --nvfp4 | |---|---|---|---| | cli.py --model distilled --no-pup | 31 GB | 26 GB | 25 GB | | cli.py --model base --no-pup | 31 GB | 29 GB | 26 GB |

The upsampler reserves --pup-memory of the GPU (default 0.35). --fp8 needs Ada, Hopper or Blackwell; --nvfp4 needs Blackwell and the [nvfp4] install, not with --offload.

This code has no int8 path. On GPUs without FP8 (RTX 30 series and older) use the int8 files from black-forest-labs/flux-3-image-klein-comfy in ComfyUI.

Safety

Off by default; --safety (CLI) or Advanced, Safety checks (demo). Prompts, reference images and outputs are checked for copyrighted characters, brands and public figures (VLM) and for NSFW content (classifier); anything flagged is refused. Saved images carry EXIF tags (Software: AI generated, Make: Black Forest Labs).

License

The code in this repository is released under the Apache License 2.0. The FLUX 3 Image Klein weights are under the FLUX Kommunity License; the prompt upsampler is Apache 2.0.

Citation

@misc{flux-3-image-klein-2026,
    author={Black Forest Labs},
    title={{FLUX 3 Image Klein}},
    year={2026},
    howpublished={\url{https://bfl.ai/models/flux-3-image-klein}},
}