P-H3-Image: Specs, License & How to Run

P-H3-Image is Pruna AI’s still-image checkpoint for text-to-image and image editing, built on MiniMax-H3. Official hosts we opened on 2026-11-04: the ungated diffusers card PrunaAI/P-H3-Image (created 2026-10-27T09:16:58Z; last modified 2026-11-03T16:08:56Z; 48 likes / 111 downloads), LICENSE (MiniMax H3 Community License Agreement, 2 August 2026), NOTICE, ComfyUI files PrunaAI/P-H3-Image-Comfy (57 likes / 52 downloads), the launch post P-H3-Image: Turning a SOTA Video Model into an Efficient Open Image Model (3 November 2026), and the Gradio Space PrunaAI/P-H3-Image (14 likes; runtime zero-a10g when opened). API gated was false. Hugging Face pipeline_tag is text-to-image. library_name is diffusers.

This identification page hangs off open Midjourney-class alternatives. That mapping is editorial. We have no sourced bake-off against Midjourney.

It is not MiniMax’s later official image checkpoint MiniMax-H3-Image (16 MP, 6 steps, up to 16 refs). Head-to-head facts: MiniMax-H3-Image vs P-H3-Image. It is not Pika H3 Flex (video+audio adapters on the same MiniMax-H3 base). Same pack versus Klein: P-H3-Image vs FLUX 3 Image Klein.

The card’s “strongest at photorealism, film and editorial looks, and graphic styles with short quoted text” sentence stays vendor.

What we verified (2026-11-04)

Field	Value	Source
Host (diffusers)	huggingface.co/PrunaAI/P-H3-Image (not gated)	HF API
Created / last modified	2026-10-27T09:16:58Z / 2026-11-03T16:08:56Z	same
Likes / downloads	48 / 111	same
License tag	other / minimax-h3-community-license-agreement	HF cardData
LICENSE file	MiniMax H3 Community License Agreement, 2 August 2026	LICENSE
Pipeline tag	text-to-image	HF API
Base model	MiniMaxAI/MiniMax-H3 (finetune)	HF tags
Transformer class	MiniMaxH3Transformer3DModel (50 layers, 56 heads)	transformer/config.json
Params (HF)	33.1B; safetensors 33,122,992,896 (BF16 33,105,770,752 + F32 17,222,144)	HF API / card
usedStorage (diffusers)	186,844,716,806 bytes (~174 GiB)	HF API
Sampling	8 steps, no CFG or negative prompt	HF README
Resolution	Best at 8 MP (default 2912 × 2912); edits 4 or 8 MP	HF README
Editing	Up to 8 input images (Picture 1 is edited; the rest are references)	HF README
Prompt upsampler	Qwen3.5-9B, on by default	HF README + NOTICE
ComfyUI host	PrunaAI/P-H3-Image-Comfy; 57 likes / 52 downloads	HF API
Official Space likes	14	HF Spaces API

The 3 November 2026 blog also says the transformer backbone “goes from 33B to 20B parameters.” We do not treat that as a second identification count: the HF card and API still print 33.1B. The same blog’s intro says “up to 9 input images”; the HF and Comfy cards we opened say 8. Identification follows the cards.

Copied 8 MP sizes from the card (width × height): 1:1 2912×2912; 4:3 3360×2496; 3:4 2496×3360; 3:2 3552×2368; 2:3 2368×3552; 16:9 3872×2176; 9:16 2176×3872.

How to run (from official cards, not tested)

Diffusers card: install git+https://github.com/huggingface/diffusers, then ModularPipeline.from_pretrained("PrunaAI/P-H3-Image"). Load bf16, move to CUDA. Pass prompt for text-to-image; pass image (one image or a list) to edit. The pipeline picks the task from its inputs. Copied sketch:

from diffusers import ModularPipeline
import torch

pipe = ModularPipeline.from_pretrained("PrunaAI/P-H3-Image")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")
image = pipe(prompt=prompt, height=2912, width=2912, output="images")[0]

At 8 MP in bf16 the card names about 101 GB with the upsampler (one H200 or B200), or 80 GB with offloading. forkalt has not run this.

ComfyUI card: native nodes, no custom nodes. Support is on ComfyUI master / nightly until the next stable release. Download one file per component (bf16, int8, or w4a8; mixed precision is allowed). Workflows in workflows/ render at 8 MP in 8 steps through the prompt upsampler. A 2 GB LoRA (loras/p-h3-image_lora_bf16.safetensors) runs P-H3-Image on an already-downloaded MiniMax-H3 FL2VA transformer. Those are vendor recipes.

The Hugging Face Space is ZeroGPU-hosted. Queue and daily GPU allowance are Hugging Face product limits, not the weight grant.

Vendor speed and memory (self-reported)

Not a forkalt measurement. Copied labels only.

Diffusers README, one NVIDIA B200, bf16, batch 1, 8 MP, no compilation or offloading: text-to-image with the prompt as written about 9 s / 83 GB; the prompt upsampler adds 6 to 11 s and raises peak memory to 101 GB.

3 November 2026 blog table (50-step MiniMax-H3 vs 8-step P-H3-Image, “without further optimizations”): 4 MP on B200 12.4 s → 2.6 s; 8 MP on B200 30.6 s → 5.8 s; 4 MP on L40S (48 GB) 19 s (bf16, offloaded) → 6.3 s (4-bit). The 8 MP 5.8 s row does not match the diffusers card’s about 9 s. We record both and do not pick a winner.

ComfyUI card peak VRAM at 8 MP: bf16 56 GB, int8 31 GB, w4a8 21 GB. Download sizes for those sets: 55 / 30 / 20 GB. The blog’s “runs in about 21 GB of VRAM for 8 MP images” is that smallest Comfy set.

License

LICENSE is the MiniMax H3 Community License Agreement (2 August 2026). Hugging Face cardData.license is other. We do not invent a more specific SPDX id.

Paraphrase only, not legal advice:

• Applicable Territory is worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States. Use outside that territory is not authorized by this agreement (contact MiniMax for a separate license).
• Yearly revenue over USD 20 million requires prior written MiniMax authorization ([email protected]).
• Prominently display “MiniMax H3” on commercial product UIs (LICENSE §IV). The HF card and blog also require “Powered by MiniMax H3”.
• You may not use the Works or Outputs to improve any other AI model other than MiniMax H3 or its Model Derivatives.

NOTICE: transformer/, vae/, text_connector/, and text_encoder/ are MiniMax H3 Community License (text encoder also based on Qwen3-VL-4B; those Qwen parts Apache-2.0). prompt_upsampler/ is based on Qwen3.5-9B under Apache License 2.0 (LICENSE-APACHE). The 3 November 2026 blog adds that ComfyUI nodes are MIT-licensed; that sentence is blog-only (not on the Comfy card LICENSE we opened).

What we did not invent

• Photorealism / film / editorial / short-quoted-text strengths stay vendor.
• “7× smaller / 5× faster than H3” and the B200 / L40S timing tables stay vendor. We do not reconcile the card’s ~9 s with the blog’s 5.8 s.
• Peak VRAM is not independently measured.
• We do not rank P-H3-Image against MiniMax-H3-Image, FLUX 3 Image Klein, Qwen-Image-2.1-Turbo, MOSS-Image, Iris-3B, or Midjourney.
• The blog’s “up to 9 input images” does not override the cards’ 8.

Related on this site: MiniMax-H3-Image — MiniMax-H3-Image vs P-H3-Image — FLUX 3 Image Klein — P-H3-Image vs FLUX 3 Image Klein — Qwen-Image-2.1-Turbo — MOSS-Image — Iris-3B — Fibo 2.0 (CC BY-NC) — Pika H3 Flex (same MiniMax-H3 base, video) — ComfyUI. Hub: Midjourney-class alternatives. Catalog: home. Corrections: contact.