MiniMax open-sources H3 Image for both image generation and editing
11/19/2026 — 11/22, 08:17·2 sources (all)·2 reports (all)
Key takeaways
- MiniMax released and open-sourced H3 Image, with weights available on Hugging Face and GitHub. Built on the H3 video model, the 33.1B-parame…
Background & analysis
On November 10, MiniMax released the image model H3-Image and simultaneously open-sourced its weights on Hugging Face and GitHub. Built on the H3 video model, H3-Image handles both image generation and image editing in a single model. Editing accepts up to 16 reference images, referenced in the prompt as Image 1, Image 2. At 4K it reaches roughly 16 million pixels, and a 21:9 aspect ratio can be produced at 6272 × 2688. Output images can also be fed directly as the first frame into H3 Max to produce video. The model carries 33.1B parameters, with weights at 69 GB plus a 67 GB text encoder, and the official hardware requirements point to the H200 and RTX PRO 6000 tier. The United States, the European Union, the United Kingdom, and South Korea fall outside the royalty-free licensing scope.
On November 21, TMTPost published a hands-on comparison of editing capability, pitting MiniMax's H3-Image, ByteDance's Seedream 5.5 released on November 9 (accessed through Doubao's AI creation entry), and Alibaba's earlier Qwen-Image-3.0-Pro (accessed through the Qwen AI platform) against one another. Using the same scene prompt, a girl in a red raincoat standing at the entrance of a convenience store on a rainy night holding a bag of milk, the author generated images in five styles: realistic photography, children's illustration, blind box figure, cyberpunk, and flat illustration, then ran three rounds of continuous local edits. The point of the test was whether each model could modify only the requested region instead of redrawing the whole image, and two of the rounds asked for the specified text "便利店 Convenience Store" on the picture. The results were mixed: H3-Image stayed relatively restrained, handling the realistic shot naturally while keeping the illustration simple, with part of the store cropped away; Doubao focused on the figure with a blurred background and richer environmental rendering; Qwen-Image-3.0-Pro produced a vertical composition with a noticeable traditional Chinese flavor and finely drawn details.
AI-generated from 2 reports · updated 3 hours ago
Latest turnThe author compared ByteDance's Seedream 5.5 on Doubao, MiniMax's H3-Image and Alibaba's Qwen-Image-3.0-Pro using the same prompts to generate five visual styles, then applied three consecutive local edits. Each model showed different tendencies: H3 stayed relatively disciplined, Doubao favored richer scenes, and Qwen sometimes produced unexpected results. The test focused on whether each model could modify only the requested part without redrawing the whole image.
Related AI tools
- Heat index
- 54
- All sources
- 2
- All reports
- 2
- First seen
- 3 days ago
Reports on this story headlines open the original
The author compared ByteDance's Seedream 5.5 on Doubao, MiniMax's H3-Image and Alibaba's Qwen-Image-3.0-Pro using the same prompts to generate five visual styles, then applied three consecutive local edits. Each model showed different tendencies: H3 stayed relatively disciplined, Doubao favored richer scenes, and Qwen sometimes produced unexpected results. The test focused on whether each model could modify only the requested part without redrawing the whole image.
钛媒体AI score 54
MiniMax has released and open-sourced H3 Image, a single model that handles both image generation and editing with up to 16 reference images; weights are on Hugging Face and GitHub. Built on the H3 video model, its output can serve as a first frame for H3 Max video, but at 33.1B parameters it needs H200-class hardware and carries royalty restrictions in the US, EU, UK and South Korea.
V2EXAI score 75
Other stories people are talking about
- Heat index 596RisingThinking Machines Lab Open-Sources Inkling 2 Mini with 32GB Model Card Recommendation10 sources
- Heat index 85Together AI offers Qwen4 Max at 30% off, fixing a Tornado bug in one request1 sources
- Heat index 38Together AI launches Qwen3 TTS with 9 voices across 10 languages1 sources
- Heat index 78LookAtMyQuant open-sources compiler that quantizes Qwen3.5-4B to 1.23 GB1 sources
- Heat index 1143RisingAnthropic Reaches $1.6 Trillion Market Value in First Trading Week After IPO21 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion

