FLUX.1 Dev FP8 v1.0
FP8 checkpoint of Black Forest Labs' 12B FLUX.1 [dev] text-to-image model, stored at one byte per weight instead of two. A run returns up to eight PNG images at up to 2048 pixels per side.
Overview
FLUX.1 Dev FP8 v1.0 is an FP8 checkpoint of FLUX.1 [dev], Black Forest Labs' 12-billion-parameter rectified flow transformer for text-to-image. Its weights are stored as float8_e4m3fn, one byte per weight instead of two in the original BF16 release, so the file is about half the size. The uploader's notes point to Kijai's FP8 release as the source of the file.
Only the transformer is quantized. The text encoders (CLIP-L and T5-XXL) and the autoencoder are the standard FLUX.1 [dev] components.
What you can build
- Cinematic concept frames. The samples include a T-rex facing a man in a yellow raincoat in a dark, misty forest, raptors riding motorcycles, and a cowboy riding a T-rex among triceratops.
- Portraits, such as a close-up against a dark background with warm side light and visible skin texture.
- Creature design, like the twig-limbed forest creature with glowing orange eyes.
- Short text inside a scene. One sample shows a robot holding signs that read "We will be back soon." and "Maintenance Mode".
- Mood board batches, since one run returns up to eight images from the same prompt.
Inputs
- A text prompt, required. The samples use full sentences with camera terms such as "cinematic medium shot".
- Width and height up to 2048 pixels each, default 1024 by 1024.
- Inference steps, default 4.
- Guidance scale from 1 to 20, default 3.5.
- Number of images, 1 to 8.
- Seed, default 123456. A seed of 0 picks a random one; any other seed with the same settings gives the same image.
There is no negative prompt field, so describe what you want to see directly in the prompt.
Outputs
One PNG file per requested image, at the width and height you set when both are multiples of 16.
Recommended settings
- The default of 4 steps gives a soft draft with little fine detail. The samples on this page used 10 to 30 steps, mostly 20 to 25. Black Forest Labs' reference example uses 50.
- The default guidance scale of 3.5 matches Black Forest Labs' reference example.
- Keep width and height at multiples of 16, for example 768 by 1360 for a tall frame or 1360 by 768 for a wide one.
Limitations
- Rounding the weights to 8 bits means results will not match the BF16 FLUX.1 [dev] pixel for pixel at the same seed and settings.
- The CLIP-L encoder reads only the first 77 tokens of a prompt. In long prompts, put the subject and the main style words first.
- Black Forest Labs notes that the model can miss parts of a prompt, depends heavily on prompting style and can amplify societal biases. It is not a source of facts.
Safety & compliance
This FP8 checkpoint is a derivative of FLUX.1 [dev], so it falls under the FLUX.1 [dev] Non-Commercial License, and the license's restrictions on the model apply to it as well. Black Forest Labs claims no ownership of outputs, and the license allows using them for any purpose, commercial use included, except where it says otherwise. Outputs may not be used to train, fine-tune or distill a model that competes with FLUX.1 [dev], and they must be disclosed as AI-generated where the law requires it. Black Forest Labs also rules out content that harms minors, non-consensual nudity, harassment and disinformation meant to harm others.
Example prompts
Great starting points for FLUX.1-dev-fp8-v1.0.
API quick start
Run FLUX.1-dev-fp8-v1.0 with a single API call.
{
"prompt": "Cute furry feral pony in a rural peaceful…",
"steps": 4,
"scale": 3.5,
"samples": 1
}curl -X POST "https://api.wiro.ai/v1/Run/abdallahalswa80/flux-1-dev-fp8-v1-0" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Cute furry feral pony in a rural peaceful…",
"steps": 4,
"scale": 3.5,
"samples": 1
}
JSONOther models you might like
Explore similar image generation models.



