Pruna AI · Open image model

P-H3-Image

Pruna AI has released P-H3-Image, an open-weights model for text-to-image generation and image editing, built on the MiniMax-H3 video model. It renders images from 2 to 16 megapixels and edits with up to 9 input images. Pruna reports it is 7x smaller than H3 (20 GB instead of 143 GB in the smallest version) and 5x faster (2.6 s instead of 12.4 s for a 4 MP image on one B200). The 4-bit version runs in about 21 GB of VRAM for 8 MP images. Pruna reworked the prompt upsampler (a small Qwen3.5-9B-based model), the text encoder (distilled from 32B to 4B), the diffusion schedule (50 steps cut to 8 with DMD distillation), the transformer backbone (33B to 20B parameters) and the decoder, then quantised to 8-bit and 4-bit. Pruna says the 8-step model matches the 50-step original for most prompts, and lists weaknesses with very short or overly long prompts, hands, fine details and small or dense text. It is released on 3 Nov 2026 under the MiniMax H3 Community License, which excludes the EU, UK, South Korea and US unless separately licensed; the Qwen-based parts are Apache 2.0 and the ComfyUI nodes are MIT. It is available as a Modular Diffusers pipeline and as ComfyUI weights.