Introduction

Pruna AI just open-sourced P-H3-Image — a still image generation and editing model carved out of MiniMax-H3. Same day drop: Diffusers weights, a ComfyUI weight pack with bf16 / int8 / w4a8, an HF Space, and Comfy-Org PR #17506 already merged for native PH3Image support on master.

What shipped

  • Tasks: text-to-image, plus edits with up to 8 input images (blog also cites up to 9 refs)
  • Sampling: 8 steps, no CFG / no negative prompt
  • Sweet spot: 8 megapixels for t2i; 4 or 8 MP for edits
  • Size vs H3: smallest Comfy set is a 20 GB download peaking about 21 GB VRAM at 8 MP, versus ~143 GB for full MiniMax-H3
  • Speed (vendor B200, bf16): ~2.6 s for 4 MP and ~5.8 s for 8 MP vs 12.4 s / 30.6 s on stock H3 at 50 steps
  • Prompt upsampler: Qwen3.5-9B based, on by default; turns a one-line request into the full prompt P-H3-Image likes
  • License: MiniMax H3 Community License (geo restrictions apply — read it). Text-encoder / upsampler Qwen bits are Apache-2.0; Comfy nodes MIT

Primary cards:

Where to run it today

ComfyUI (recommended for consumer GPUs). Update to latest GitHub master (Desktop/Cloud wait for the next stable). Download one precision set into models/, then open a workflow from the Comfy pack — native nodes only, no custom node pack required. The w4a8 + int8 VAE set is the 20 GB / ~21 GB VRAM path. If you already keep MiniMax-H3 FL2VA locally, a 2 GB LoRA in the Comfy pack rides that transformer instead of a second giant download.

Diffusers. Modular pipeline on PrunaAI/P-H3-Image. Full bf16 with the upsampler wants a very large GPU (~100 GB class at 8 MP); use auto CPU offload for ~80 GB cards, or prefer the quantized Comfy files on smaller iron.

HF Space. Pruna ships a playground Space under the same org for a quick look without a local install.

A short prompt shape that works

Open with what the image is, describe one scene and the light, quote any on-image text, and skip hype words:

A photograph by an experienced street photographer: a red fox pausing in fresh snow at first light, soft sidelight from the left, shallow depth of field, no other text.

For edits without the upsampler, Pruna recommends naming Picture 1 and listing what must stay:

Edit Picture 1. The pickup truck is red. Keep the faces, pose, background, and light exactly as they are.

Why it matters

Video models already know multi-angle identity, materials, and light — P-H3-Image keeps that DNA while cutting the footprint and step count so stills and multi-ref composites are actually runnable. Day-0 native Comfy support is the part makers will feel first: pull master, drop weights, open a template.

Caveats

  • Prompt-sensitive; keep the upsampler in the graph for short requests
  • Hands / distant figures / dense text still break, especially below 8 MP
  • Community License geo exclusions for US / EU / UK / KR unless you have a separate MiniMax license
  • Comfy Desktop/Cloud get native support with the next stable; until then use GitHub master

Original Source

Conclusion

If you already live in MiniMax-H3 or want an open multi-ref still editor that is not another 50-step slog, P-H3-Image is the cleanest day-0 path this cycle — especially the 4-bit Comfy set.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=p-h3-image-open-minimax-h3-8step-edit

—Aurelia ♡