Open Iris-3B Turbo Generates 1024px Images in 0.7s on One H100 With Just 4 Steps
lmoroney · x · 2026-10-15
Speridlabs shipped Iris-3B Turbo, a 4-step distilled version of its open 3B pixel-space text-to-image model (Apache 2.0, weights on Hugging Face).
- Speed: 0.7s for a 1024×1024 image on one H100, fixed 4 steps vs. the base model's default 100.
- Architecture: pixel-space diffusion transformer that skips the VAE used by Flux/Qwen-Image/SDXL, with a frozen Qwen3-VL-4B-Instruct text encoder; trained with rectified flow at 256→512→1024px over 665K steps.
- Turbo: full-parameter data-free distillation via Trajectory Distribution Matching, CFG baked into weights.
- Same backbone also ships as a monocular depth model and 4× restoration/upscaler.
More from Multimodal
- Reverie ships Intercut Clips to fix AI video's hardest shot: dialogue scenes — lmoroney · 2026-10-15
- Midjourney sref code produces glitter-textured Pokémon starter illustrations — sergeantsref · 2026-10-15
- TRELLIS.2 image-to-3D runs at full 1536 resolution on a 16 GB card via tiling patches — over-yonder-tech · 2026-10-15
- OpenSpline: open-source avatar layer adds a realtime lipsynced face to your voice agent in 5 lines — Mikasa0xdev · 2026-10-15
- Three months of Claude-guided DiT fine-tuning fails, non-technical dev begs Reddit for help — Old_Manufacturer7905 · 2026-10-15
- FastH3-Omni Launch: Real-Time Pre-Viz Studio Built via Vibe Coding, Open Source — Scobleizer · 2026-10-15