3B-parameter model Iris generates every pixel directly, no VAE needed
Deadity · reddit · 2026-10-10
A Reddit user shared speridlabs/iris-3b on Hugging Face: a 3-billion-parameter image generation model that skips the VAE entirely and models every pixel directly. It's an unusual architecture choice versus mainstream latent-diffusion approaches, and an interesting study for anyone exploring non-diffusion generative routes.
Related event: Sperid Labs Open-Sources Iris-3B, a Pixel-Space Text-to-Image Model(8 posts)→
More from Multimodal
- Step 5 Preview generates a 30-second motion clip in one shot via Hermes agent — Teknium · 2026-10-11
- Tencent's open-source Hunyuan3D-2 turns one image into textured 3D models, 15k stars — Promptmethus · 2026-10-11
- Open-source music model YuE2 runs locally on a MacBook Pro, rivaling Suno quality — vista8 · 2026-10-11
- Two prompts + Claude made a Monet-style MV for Jay Chou's 'Qi Li Xiang' — AlchainHust · 2026-10-11
- Underdog launches on-device image gen powered by Qwen models, photos never leave your computer — Scobleizer · 2026-10-11
- Open-source anime DiT models: Anima staleness has users pinning hopes on Krea2 fine-tunes — Pristine_Stress_670 · 2026-10-11