Open-Source Realism LoRA for MiniMax H3: Curing the Plastic Look in AI Video
Affectionate-Map1163 · reddit · 2026-08-11
To tackle the lack of realism in AI-generated video humans, a developer trained and open-sourced Realism People, a LoRA adapter for the MiniMax H3 video model.
Key Improvements:
- Maintains natural skin texture, eliminating the typical "plastic" AI look.
- Keeps eyes and micro-expressions coherent, with lighting behaving like a professional film set.
- Adds a subtle handheld, documentary feel to the motion while preserving H3's native synchronized audio.
Training & Selection Details:
- Tested 16 different configurations across two dataset versions.
- The winner was selected through 100 same-seed A/B duels (same prompt and seed, adapter on vs. off).
- Winning config: Rank 16, trained for 5,000 steps at a low learning rate.
Usage:
- Weights are open-source on Hugging Face. Use the trigger word r34l1sm at the start of your prompt.
- Intended strength is 1.0 (scale down to 0.6-0.8 for a lighter touch).
- Works with H3's LoRA endpoints: text-to-video, image-to-video, and reference-to-video.
Related event: Open Source Realism LoRA Enhances MiniMax H3 Video Characters(4 posts)→
More from Multimodal
- Wan 2.2 VACE vs Minimax H3: Which Model is Better for Video Edit Consistency? — equanimous11 · 2026-08-11
- ComfyUI Tip: Fast and Accurate Upscaling with Nvidia RTX Nodes — R4ndy0rt0n · 2026-08-11
- Real-Time Mocap Breakthrough: Capturing Bodies, Fingers, Faces, and Multi-Character — andrew_n_carr · 2026-08-11
- Testing a New AI Motion Capture Workflow: 30-Second One-Shot Generation — Chuka444 · 2026-08-11
- First 110-Minute AI Feature Film Made for $2M Open-Sources All Prompts — dr_cintas · 2026-08-11
- MiniMax H3 Video LoRA Training Fails: Unable to Capture Identity — LordDarthShader · 2026-08-11