MiniMax H3 Image
by MiniMax · Model · Image
One of the best open image models, fast and close to Qwen Image 2.1
Open image model generating 4K images in seconds with natural-language editing and reference images; called one of the best open image generators
Why we say so
Artificial Analysis ranks it first among open-weights models for editing and text-to-image, and Arena puts it second for open editing (1358, just behind Qwen Image 2.1's 1363) and third for open text-to-image. AI Search found its edits very close to Qwen Image 2.1, but Krea 2 and MOSS-Image beat it on some realism prompts, and it does poorly at 1 megapixel. Fahd Mirza saw garbled small text and failures on complex scenes, and reported VRAM use from 16 GB (AI Search) to nearly 48 GB, against MiniMax's 6 GB claim.
Ranks #2 for image editing (open weights) · #3 for image generation (open weights)
Evidence
- Nov 17#2 on Arena image-edit among open weights (1358)
- Nov 17#3 on Arena text-to-image among open weights (1198)
- Nov 15Open-weight 33B diffusion transformer distilled to 6 sampling steps, no CFG, native ~16 MP output without upscaler
- Nov 14
- Nov 12
Show all 8
- Nov 12Edit quality very close to Qwen Image 2.1; better at fixing lighting and colour, less faithful when restoring old photos
- Nov 12Krea 2 better at realism, yoga pose anatomy, dot illustration and Monet style; MiniMax wins UI design, magazine cover, soles prompt
- Nov 12Reviewer thinks Krea 2 or MOSS-Image are still better for text-to-image in some respects, but MiniMax is far faster at 2K/4K