[Comparison] MOSS-Image-SFT vs MiniMax-H3-Image(Pruned): A Fair 1:1 Benchmark Across 10 Diverse Art Mediums (Identical Prompts, Zero Negative, 1K Resolution)
retained text
Hi everyone, I put together a head-to-head baseline comparison between two major new architectures: **MOSS-Image-SFT** and **MiniMax-H3-Image**. To keep the comparison 100% fair and avoid fine-tuning bias: * **100% Identical Prompts**: Both models received the exact same descriptive visual prose describing the medium, anatomy, lighting, and composition. * **Zero Negative Prompt**: No negative conditioning was used (`negative_prompt = ""` on MOSS, matching H3's native CFG 1.0 architecture). * **Pure Original Characters & Techniques**: Zero proper names and zero artist name-dropping. Styles were guided purely through physical medium mechanics (lithograph, Shin-Hanga woodblock, 90s analog cel, Kirie papercut, Guohua ink splash, etc.). * **Strict 1:1 Aspect Ratio & 1K Resolution**: Exact matching seeds and native \~1.0 MP resolutions (2:3, 16:9, 1:1, 4:3, 4:5, 9:16, 3:2, 3:4). * There is version for specifict finetuned for the two model, but the result is similar to this. * I notice Minimax-h3-Image struggling on generating limbs. Every prompt i used the model always have some error. * A Fully finetuned version for both models, set at 4K (minimax-h3-image) and 2K (moss-image-sft). Prompt workflows included : [https://drive.google.com/drive/folders/1veLGRIMtd-o4zN3mS160uHlOLUuf8vgt?usp=sharing](https://drive.google.com/drive/folders/1veLGRIMtd-o4zN3mS160uHlOLUuf8vgt?usp=sharing) Workflows (standard) : [Moss-Image-SFT WF](https://pastebin.com/U1YDKsC3) [Minimax-H3-Image WF](https://pastebin.com/aExAgN3G)
document fetched 2026-11-12 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xtjm4z/comparison_mossimagesft_vs_minimaxh3imagepruned_a