HeyGen trains avatar behaviors as separate LoRAs, then composes them at inference
HeyGen · x · 2026-07-25
HeyGen says it can control multiple avatar behaviors without a jointly labeled dataset by training each behavior — expressiveness, gaze, and camera — as separate LoRAs on single-attribute data.
At inference time, the behaviors are composed together, and the train/inference gap is narrowed with a short fine-tune on real overlapping samples.
More from Multimodal
- Side-by-side aircraft render comparison shows Opus 5 as the more detailed version — victormustar · 2026-07-25
- FLUX 3 video pretraining is being pitched as a boost for robot learning — rohanpaul_ai · 2026-07-25
- Opus 4.1’s art-session energy becomes a shareable AI joke — liminal_bardo · 2026-07-25
- SuperSplat Leverages WebGPU and Voxel Collision for Massive Point Cloud Rendering — willeastcott · 2026-07-25
- HeyGen shows video background removal with transparent foreground, mask and plate — HeyGen · 2026-07-25
- Google says agent behavior comes from prompts, evals, iteration, and feedback loops — AI Engineer · 2026-07-25