bluzen/prismstream-xl

prismstream-xl

Flow-matching text-to-image with region conditioning: a mask is a statement about composition, not a repair instruction. Trained with masks in the conditioning path from step one rather than bolted on afterwards.

How to use it

from diffusers import PrismStreamPipeline

pipe = PrismStreamPipeline.from_pretrained("bluzen/prismstream-xl")
image = pipe(
    prompt="a quiet street at dusk",
    regions=[{"mask": mask, "prompt": "a red umbrella"}],
    num_inference_steps=12,
).images[0]

Twelve steps is the quality we publish numbers for. Four is usable for previews and visibly worse.

Limitations

• The API is not stable. This is the fastest-moving project in the lab.
• Region boundaries degrade when two regions overlap by more than about 30%.
• Text rendering inside an image is bad, in the ordinary way.