Novice tests 6 video models for digital-human training videos, none passes
Warshawski001 · reddit · 2026-11-20
A company tech lead, camera-shy and with unclear voice recordings, tested multiple video models in ComfyUI to generate 5-second digital-human clips for software training videos. Results: Veo 3.1 Lite added gestures and stuttered dialogue; Omni1.2 Flash was best but changed the scenery; MiniMax H3 kept scenery but produced nonsense dialogue; Pika H3 Flex looked green-screened with shrinking artifacts; Wan 2.2 walked weirdly and hopped. His conclusion: no current model preserves face, scene, and dialogue simultaneously, and he's weighing stills vs. real footage with character replacement — a snapshot of reference-to-video's real-world limits.
More from Multimodal
- Cinematographer Tests Minimax's Cinestill 800T Emulation for a Blade Runner Look in Neon Rain — Purple-Let4907 · 2026-11-21
- 15-year electronic producer built an agentic DAW with Claude and made a full album — yawnxyz · 2026-11-21
- Self-hosted cloud GPUs beat Higgsfield and Runway APIs on cost and flexibility — Al_Grigor · 2026-11-21
- Gaussian splat top-down scan shows how Fable plans object placement in pixel apartment — XRarchitect · 2026-11-21
- Pixel replica of a real apartment built from Atlas gaussian splats with Fable 5.2 — XRarchitect · 2026-11-21
- Fable 5.2 autonomously drafts detailed architectural plan of John Hancock building in Aircad — iskander · 2026-11-21