Repository: localaiLicense: apache-2.0
Iris 3B is a text-to-image diffusion model using a Qwen3-VL-4B-Instruct text encoder. Its Iris diffusion weights are the original F32 safetensors (~11.1 GiB), while the encoder is Q4_K_M (~2.3 GiB). Iris generates directly in pixel space, so it does not use a VAE. Use widths and heights divisible by 16; output around one megapixel is recommended.
Links
Tags