Skip to content

Generate images with diffusion models

mistral.rs generates images from text prompts with FLUX.1, served at POST /v1/images/generations.

Terminal window
mistralrs serve -m black-forest-labs/FLUX.1-schnell

There are two checkpoints:

Model Denoising steps Prompt length Access
black-forest-labs/FLUX.1-schnell 4, so it’s fast Up to 256 tokens. Longer prompts return an error. Open
black-forest-labs/FLUX.1-dev 50, with guidance scale 4.0 Not limited to 256 tokens Gated: accept the license on its Hugging Face page, then run hf auth login or set HF_TOKEN

The step count is fixed per checkpoint; request fields like steps and guidance_scale are ignored.

Terminal window
curl http://localhost:1234/v1/images/generations \
-H 'Content-Type: application/json' \
-d '{
"model": "default",
"prompt": "A golden retriever wearing a scarf in autumn leaves.",
"height": 1024,
"width": 1024,
"response_format": "B64Json"
}'
Field Default Meaning
prompt Required What to draw
n 1 How many images to generate
height, width 720, 1280 Output size in pixels. OpenAI’s size string (such as "1024x1024") is ignored, so set these instead.
response_format Url Url saves the image on the server and returns its filename. B64Json returns the image inline. See Output handling.

FLUX can’t be quantized: --quant and --isq don’t apply to it. The FLUX transformer alone has about 12B parameters, roughly 24 GB at 16-bit precision, and the T5-XXL text encoder needs more memory on top while it reads the prompt.

If that doesn’t fit on your GPU, load the offloaded variant from Python or Rust (FluxOffloaded, as in the examples above). It keeps the model in system RAM and moves each layer onto the GPU only while that layer runs. GPU memory use drops sharply, but every step copies the model across, so images take longer. The CLI and server always load the full model onto the GPU.

With B64Json, b64_json holds a complete data URL (data:image/png;base64,...). Strip the prefix before decoding:

import base64
payload = response.data[0].b64_json.split(",", 1)[1]
with open("out.png", "wb") as output:
output.write(base64.b64decode(payload))

With Url, the server writes a PNG named like image-generation-<uuid>.png into its working directory and returns that filename in data[0].url. It’s a path on the server’s disk, not a download link, so use B64Json when your client runs on another machine. In the SDKs, pass save_file to choose where the PNG goes.