Iris 2 pixel model lands in ComfyUI with custom nodes, showing VAEs aren't mandatory
Incognit0ErgoSum · reddit · 2026-10-10
A developer wired the Iris 2 pixel model into ComfyUI (envy-ai/ComfyUI-Iris), shipping custom loader and sampler nodes since pixel models work differently from VAE-based ones. LoRAs work as usual and ComfyUI's native schedulers/samplers are supported. Results fall short of Anima 2.9B at similar size, but it demonstrates VAEs aren't a necessary evil. Recommended: dpmpp-2m-sde-gpu sampler; bongtangent scheduler helps. A forked ai-toolkit repo can train LoRAs for it.
More from Multimodal
- Creator shares full AI ad workflow: no editor, no After Effects, all free tools — faraz0x · 2026-10-10
- AI video outtakes turned into a blooper reel: humor or completely wacko? — Kyrannio · 2026-10-10
- Mutuals to open-source real-time 3D character harness built with Oscar-winning animators — gdechichi · 2026-10-10
- Open-source ComfyUI node pack adds reusable image/audio/video references for MiniMax H3 RefMods — Entert-AI · 2026-10-10
- Alibaba's TaoMate-H3 streams minute-long video with audio using 3-step denoising — jiqizhixin · 2026-10-10
- SGLang-Diffusion serving framework for diffusion models to be unveiled at PyTorch Conference 2026 — PyTorch · 2026-10-10