From a ComfyUI character LoRA to an 18-second action sequence with Seedance
alecubudulecu · reddit · 2026-11-01
A Reddit creator details a full pipeline from still character to an 18-second action sequence: train a character LoRA on Z-Image Turbo, render references in ComfyUI, then feed those refs into Seedance 2.5 to generate four short shots, edited into one sequence.
The key lesson is shot-by-shot generation: splitting the scene into short clips lets you re-roll weak motion individually without losing the whole scene, balancing character consistency with controllability. The author offers to share references and workflow details.
More from Multimodal
- UniT framework trains unified multimodal models to plan, inspect, and refine iteratively — burkov · 2026-11-01
- PlotWeaver unveils STORM-Ensemble 3.0: one 3B model synthesizing 64 African languages — Fowe · 2026-11-01
- World models are 'horrible' at accurate scenes; dev sticks with Gaussian splatting — WesEklund · 2026-11-01
- AI-made music video 'UpRyZ' mashes 80s zombie horror with rock and dance moves — lump- · 2026-11-01
- Architect praises AI-generated isometric city maps, 620 users exploring simultaneously — miilesus · 2026-11-01
- Redditor recreates a viral K-beauty video with AI — can you spot the real one? — Zero316 · 2026-11-01