MiniMax H3 Anime Video Test: Impressive Face Consistency but Prone to Motion Jitter
OneTrueTreasure · reddit · 2026-08-01
A Reddit user tested MiniMax H3's Reference-to-Video (R2V) capabilities using anime fight scenes. Strengths: It outperforms Wan 2.2 and LTX 2.3 in anime generation, shows excellent facial consistency, and its motion control successfully maintains 2D anime aesthetics without making them look 3D.
Weaknesses: The current API output suffers from noticeable motion artifacts and jitter, with Text-to-Video (T2V) appearing blurry and undercooked. The author speculates this might be due to suboptimal API default settings (like CFG or steps) or low-resolution reference inputs.
Future Outlook: The author hopes that once open-weights are released tomorrow, the community can improve fluidity using LoRAs, fine-tuning, or distillation.
Related event: MiniMax H3 API Tests Reveal High Quality and No Censorship(3 posts)→
More from Multimodal
- ByteDance's Seedance 2.5 release leaked; open-source project exposes API — matchaman11 · 2026-08-01
- Using LichtFeld to Transform DJI 8K 360 Video into 3D Gaussian Splatting — janusch_patas · 2026-08-01
- Open-Source Video Model H3 Praised for SOTA Motion Physics and Uncensored Access — Neggy5 · 2026-08-01
- ComfyUI v0.28 Update Apparently Causes Complex Widgets to Vanish in Subgraphs — ChrisUnimportant · 2026-08-01
- Open-Source Krea2 Node: Bounding Boxes Solve Multi-Character LoRA Bleeding — tekprodfx16 · 2026-08-01
- AI Agents Master Houdini via MCP to Automate Procedural VFX Pipelines — anselm · 2026-08-01