Redrob Image
Open sourceA light text-to-image model that finishes in about 8 steps. The weights are Apache-2.0, so you can take them, change them, and ship them again.
Turbo sampling finishes a frame in about 8 steps instead of the 30 to 50 a normal diffusion model needs. On the same card, that is the difference between waiting and iterating.
No custom runtime, no bespoke node graph, no wrapper of ours to learn. pip install diffusers, from_pretrained, and the code you already have keeps working.
Fewer steps and a single merged UNET mean less memory and less time per image. It is meant for one consumer card, not a rented cluster.
Not open-ish. The weights themselves are Apache-2.0 on Hugging Face, so you can fine-tune them, put them in a product, and redistribute them without asking us.
The tuning went into photo, portrait, and mood instead of trying to be even everywhere. Ask it for a lit scene, a face, or a palette and it holds up.
It is weak at legible text. Hangul and most non-Latin scripts come out wrong, and typography-heavy graphic work lands well below the photo and portrait results.
The preference study against GPT Image 2 came out 44-16-30 over 90 blind pairs. The design, the confidence intervals, the categories it lost, and what we would size differently next time are all written up outside this page.
Read the measurement writeuppip install -U torch transformers accelerate safetensors && pip install -U diffusersLicense, limits, and how the path actually runs. Claims travel with method and caveat.