MiniMax-H3-Image vs P-H3-Image
Compare MiniMax-H3-Image and P-H3-Image.
| Dimension | MiniMax-H3-Image | P-H3-Image | Source |
|---|---|---|---|
| License file | MiniMax-H3-Image Community License (10 November 2026) | MiniMax H3 Community License (2 August 2026) | LICENSE files (official 2026-11-11 / P-H3 2026-11-04) |
| Excluded territories | EU, UK, Republic of Korea, United States | EU, UK, Republic of Korea, United States | LICENSE I.5 on both files |
| Commercial authorization trigger | USD 1M annualized revenue or any Hosted Service to Third Parties | Yearly revenue over USD 20 million | Official LICENSE §IV / P-H3 LICENSE §IV |
| UI attribution | Prominently display “MiniMax-H3-Image” | Display “MiniMax H3” / “Powered by MiniMax H3” | Official LICENSE §IV.2 / P-H3 LICENSE + card |
| Developer | MiniMax (official image checkpoint) | Pruna AI (still-image pack of MiniMax-H3) | Official cards |
| Hugging Face host | MiniMaxAI/MiniMax-H3-Image (not gated) | PrunaAI/P-H3-Image (not gated) | Hugging Face API |
| Parameters | 33.1B dense DiT (card); MiniMax-H3 Qwen3-VL encoder separate | 33.1B transformer (HF); 4B text encoder separate | Official cards / HF API |
| Sampling | 6 steps, guidance_scale 1, no negative prompt | 8 steps, no CFG or negative prompt | Official READMEs |
| Native resolution named | Best around 16 MP (default examples 4096²); working range ~4–20 MP | Best at 8 MP (default 2912²); edits 4 or 8 MP | Official READMEs |
| Reference images | Up to 16 (MiniMax card) | Up to 8 (P-H3 cards; blog “9” is not used) | Official cards |
| Prompt expander | Optional MiniMax-H3-Image-PE (NVFP4, 17.9 GB) or API H3-Image-Context-IR | Qwen3.5-9B upsampler, on by default in ModularPipeline | Official cards |
| Hardware (vendor) | vLLM-Omni DLO ~14.7 GiB peak at 2048²; 73.4 GiB 1-GPU offload | Comfy w4a8 ~21 GB peak at 8 MP; diffusers bf16 ~101 GB with upsampler | vLLM recipe 2026-11-11; P-H3 cards 2026-11-04 |
| Hugging Face likes | 400 (2026-11-11) | 48 on the diffusers card (2026-11-04) | Hugging Face API |
| Quality vs Midjourney | Not independently tested | Not independently tested | n/a |
When to choose which
- Official MiniMax 16 MP / 6-step image weights with up to 16 refs: MiniMax-H3-Image — License excludes EU/UK/KR/US and adds a $1M / Hosted Service trigger. Not a quality rank.
- Documented Comfy 4-bit ~21 GB peak at 8 MP: P-H3-Image — Vendor 21 GB. Same geo exclusion; older $20M H3 Community License. Not a quality rank.
- Use from the EU, UK, South Korea, or United States without a separate MiniMax license: neither — Both Community Licenses exclude those territories. Klein on this site is a different family. Not legal advice.
- Apache-2.0 still-image checkpoint: neither — Iris-3B and MOSS-Image on this site are Apache-2.0. Different size and job details.
Read the entity pages: MiniMax-H3-Image and P-H3-Image. Both are listed on Midjourney-class alternatives. Same P-H3 versus Klein: P-H3-Image vs FLUX 3 Image Klein. Local UIs: ComfyUI vs Midjourney, Fooocus vs Midjourney.
These are generation checkpoints, not pixel editors. Retouching an existing photo is PhotoCraft vs GIMP. Neither card is Pika H3 Flex (video on MiniMax-H3). MiniMax’s chat MoE on this site is MiniMax M3.1 Flash.
Why this comparison exists
MiniMax opened MiniMax-H3-Image as the official open-weight still-image checkpoint (LICENSE date 10 November 2026). This site already identifies P-H3-Image, Pruna’s still-image pack of MiniMax-H3, verified 2026-11-04. Both are few-step, no-CFG, T2I+edit transformers in the same 33.1B / 50-layer family. They are not the same artifact: different sampling, native megapixels, reference counts, text-encoder size, and license files. Identification, not a bake-off.
Sourced snapshot (MiniMax-H3-Image 2026-11-11 / P-H3-Image 2026-11-04)
| Field | MiniMax-H3-Image | P-H3-Image |
|---|---|---|
| Host | huggingface.co/MiniMaxAI/MiniMax-H3-Image (not gated) | huggingface.co/PrunaAI/P-H3-Image (not gated) |
| HF likes | 400 (2026-11-11) | 48 (2026-11-04) |
| License | MiniMax-H3-Image Community License (SPDX other) | MiniMax H3 Community License (SPDX other) |
| Size | 33.1B dense DiT (card) | 33.1B transformer (HF) |
| Sampling | 6 steps, guidance 1 | 8 steps, no CFG |
| Refs | Up to 16 (MiniMax card) | Up to 8 (P-H3 cards) |
| Native resolution named | Best around 16 MP (4096²) | Best at 8 MP (2912²) |
| Text encoder | MiniMax-H3 Qwen3-VL (~66.7 GB) | Qwen3-VL-4B (card) |
| ComfyUI | Native on 2026-11-10 Comfy blog | Native on master/nightly (card) |
| Quality vs Midjourney | Not independently tested | Not independently tested |
License
MiniMax-H3-Image LICENSE (10 November 2026) excludes the EU, UK, Republic of Korea, and United States. §V.4 also bars Outputs outside that territory. §IV requires prior written authorization if Annualized Revenue exceeds USD 1 million or you provide a Hosted Service to Third Parties, and requires displaying “MiniMax-H3-Image”.
P-H3-Image LICENSE is the older MiniMax H3 Community License (2 August 2026): same four-territory exclusion, USD 20 million yearly-revenue authorization, display “MiniMax H3” / “Powered by MiniMax H3”.
Do not mix the two files. Neither is Apache-2.0. Not legal advice.
How to choose
- Need official MiniMax 16 MP / 6-step still-image weights and can accept the November 2026 Community License: MiniMax-H3-Image.
- Need Pruna’s 8 MP / 8-step pack and the documented Comfy 4-bit path: P-H3-Image.
- Need BFL 5.6B Klein (already compared with P-H3): FLUX 3 Image Klein — P-H3-Image vs FLUX 3 Image Klein.
- Need Apache-2.0 bilingual T2I: MOSS-Image.
- Need a small Apache-2.0 pixel-space checkpoint: Iris-3B.
- Need Qwen turbo (research license): Qwen-Image-2.1-Turbo.
See also Image & Design and comparisons.
FAQ
Which one generates better images?
We have no sourced bake-off. MiniMax blog “under 10 seconds” / “5×” and P-H3 photorealism / film / editorial sentences stay vendor. Do not treat either as a Midjourney verdict.
Is P-H3-Image a packed MiniMax-H3-Image?
No. P-H3-Image is Pruna’s still-image checkpoint built on MiniMax-H3 (the video/omni model). Its diffusers card was created 2026-10-27, before this official image LICENSE date of 10 November 2026. Same 33.1B / 50-layer family, different sampling, resolution, refs, encoder size, and license file. Not a quantization of MiniMax-H3-Image.
Can I run both locally?
Both publish ungated weights. MiniMax-H3-Image needs the image repo plus the MiniMax-H3 text encoder (~68.65 GB + ~66.7 GB on the card; HF usedStorage ~190 GiB on the image card). P-H3-Image Comfy w4a8 names ~21 GB peak at 8 MP; diffusers bf16 names ~101 GB with the upsampler. None of those VRAM figures are forkalt measurements.
Which license is usable commercially?
Both exclude the EU, UK, Republic of Korea, and United States unless you have a separate MiniMax license. Official MiniMax-H3-Image adds a USD 1 million annualized-revenue or Hosted Service authorization trigger and “MiniMax-H3-Image” UI display. P-H3-Image uses the older MiniMax H3 Community License with a USD 20 million yearly-revenue trigger and “MiniMax H3” display. Neither is Apache-2.0 or MIT. Not legal advice.
Why not vs FLUX 3 Image Klein?
P-H3-Image already has that pairing: P-H3-Image vs FLUX 3 Image Klein. This page is the official MiniMax image checkpoint versus the Pruna pack already on this site.
Sources & verification
- https://huggingface.co/MiniMaxAI/MiniMax-H3-Image
- https://huggingface.co/MiniMaxAI/MiniMax-H3-Image/raw/main/LICENSE
- https://huggingface.co/MiniMaxAI/MiniMax-H3-Image-PE
- https://www.minimax.io/blog/minimax-h3-image
- https://recipes.vllm.ai/MiniMaxAI/MiniMax-H3-Image
- https://huggingface.co/PrunaAI/P-H3-Image
- https://huggingface.co/PrunaAI/P-H3-Image/raw/main/LICENSE
- https://huggingface.co/PrunaAI/P-H3-Image-Comfy
- https://www.pruna.ai/blog/oss-p-h3-image-launch
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.