Iris-3B Turbo distills pixel-space diffusion to 4 steps, 1024px image in ~0.7s on one H100
lmoroney · x · 2026-10-15
Speridlabs shipped Iris-3B Turbo, an open-weights image model. Iris-3B is a 3B-parameter diffusion transformer working directly in pixel space, skipping the VAE that Flux, Qwen-Image, and SDXL rely on. Turbo fully distills all 3B parameters down to four fixed steps with guidance baked into the weights.
ArtRealmAI reports 0.7 seconds for a 1024×1024 image on one H100 at batch 4 (base model defaults to 100 steps). On the model card, Turbo edges past the base on GenEval (0.820 vs 0.798), gives back a bit on long rendered text, and shows slightly less seed variety. Weights are Apache 2.0, 6 GB in bf16, with the distillation trainer included in the repo.
More from Multimodal
- AReA paper brings adaptive receptive-field attention to high-res diffusion transformers — Apprehensive_Sky892 · 2026-10-15
- Skip 4K drafts: one creator's math shows $5 instead of $9 per AI video clip — Div_pradeep · 2026-10-15
- 480p drafts then 4K upscale: a $5 workflow that beats $9 hi-res AI video rerolls — Div_pradeep · 2026-10-15
- A Banana Lawyer Demands Justice: Nano Banana Courtroom Video Goes Viral — NegotiationSmooth467 · 2026-10-15
- LTX 2.5 vs H3 minimax: side-by-side short-story video generation test — Numerous_Access_8008 · 2026-10-15
- Stunning AI-generated future racer tearing through a dusk cityscape — ciguleva · 2026-10-15