Exploring Wan 2.2 LoRA: How to Train Models to Keep Realistic Faces Static
Background-Tie-3664 · reddit · 2026-08-03
When generating videos with the Wan 2.2 model, the author wants to train a motion LoRA to suppress blinking or micro-expressions, keeping a realistic character's face nearly static (behavioral consistency) rather than just maintaining visual consistency.
While this is easy to achieve for anime characters, finding an effective solution for realistic human faces has proven difficult. The author notes that although videos utilizing this technique exist, the methods are gatekept within the community, prompting them to ask for concrete training advice or help.
Related event: Tackling Character Inconsistency in Wan I2V via LoRA Training(2 posts)→
More from Multimodal
- Tessellation-free geometric queries on displaced surfaces cut caustics variance 10x — ssh4net · 2026-10-16
- Complex multi-link CAD mechanism finally animates, dev celebrates the breakthrough — zetalyrae · 2026-10-16
- OpenMOSS releases MOSS-Image-SFT: Apache 2.0 text-to-image model on Hugging Face — bdsqlsz · 2026-10-16
- Hyperflow open-sourced: 49 sampling steps down to 8, ~3x faster video generation — mhdfaran · 2026-10-16
- Cinematic 3D fantasy character prompt template with salvaged-robot armor — azed_ai · 2026-10-16
- Midjourney vintage woodcut style: --sref 2446102396 beaver with a cleaver — michaelrabone · 2026-10-16