PrunaAI Open-Sources P-H3-Image: 7x Smaller, 5x Faster, 8-Step Images
sdiazlor · reddit · 2026-11-04
PrunaAI released P-H3-Image, an open-weights image model built from MiniMax-H3, available on Hugging Face and natively in ComfyUI: 5x faster generation (8 steps instead of 50), 4MP images in 2.6s vs 12.4s, 7x smaller for small GPUs. Renders 2–16MP, edits with up to 8 input images, strongest at photorealism and film stills. Every H3 component was reworked: prompt upsampler, text encoder, diffusion schedule, transformer, decoder.
More from Multimodal
- Suno song mourns AI-automated economists in Russian alt-rock — shockingly good — aniketapanjwani · 2026-11-04
- Fully Local AI Short Film Rendered on a Mac, Workflow Open-Sourced as Phosphene — Opening-Ad5541 · 2026-11-04
- Excel Shonen Anime Returns with Episode 2: 'God of Death' — superfatbeagle · 2026-11-04
- AI-generated cat 'uses up 9 lives' on a 15-second vacation in Dreamina video — umesh_ai · 2026-11-04
- YC-backed Vyra launches conversational AI agent that edits raw footage into short videos — ycombinator · 2026-11-04
- Hala Recreates Haraj's Ad Video with MCP + Claude, Generating Motion Graphics with Voiceover — aziz4ai · 2026-11-04