MOSS-Image-SFT vs MiniMax-H3-Image: Fair 10-Medium Benchmark Finds Limb Generation Failures
SkirtSwimming8950 · reddit · 2026-11-11
A Reddit user ran a strictly controlled head-to-head between MOSS-Image-SFT and MiniMax-H3-Image (pruned): identical prompts, zero negative prompts, matched seeds at 1K resolution across 10 art mediums (lithograph, Shin-Hanga, 90s cel, kirie papercut, ink splash, etc.) with no artist names. MOSS (45 steps, CFG 3.4) renders more intricate detail while MiniMax-H3 (6 steps, CFG 1.0 Flow Match) is faster but simpler. The author reports MiniMax-H3 consistently struggles with limbs. Workflows are public.
Related event: Reddit Users Benchmarks MOSS-Image Against MiniMax H3 Image(3 posts)→
More from Multimodal
- Redditor makes fast-motion short film with H3 and ships Longshot Studio v0.18.5 with face swap — R34vspec · 2026-11-12
- Reimagining Knight Rider's opening as a Netflix drama with Claude Fable 5.2 — justin_hart · 2026-11-12
- Gemini 4 Carbon builds a full 3D aquarium in Blender and Three.js in ~6 hours — sunjiao123sun_ · 2026-11-12
- Awaken Open-Sourced: Turn a Half-Duplex Speech LLM Full-Duplex With Just a Small LoRA — alexcovo_eth · 2026-11-12
- MiniMax H3 Image targets e-commerce: sketch to 4K campaign shots in ~10 seconds — MiniMax_AI · 2026-11-12
- One person can now build all of this solo with Kaigen 3D — gdechichi · 2026-11-12