Skip to content
ComfyUI Wiki · 工作流与开源创作分享·· 2026-10-18AI score58

Speridlabs releases Iris-3B pixel-space diffusion model, with CPU and Apple Silicon ports already available

Text to image

AI briefing

Speridlabs has released Iris-3B, a 3B-parameter diffusion transformer that denoises directly in RGB pixel space, with no VAE or latent space; at the end, 4 pixel heads reconstruct 16×16 patches into pixels, and the text encoder is a frozen Qwen3-VL-4B-Instruct.

Open original comfyui-wiki.com