GenCine was trained on academic compute: full technical details now on the project page
anand_bhattad · x · 2026-10-03
Closing the GenCine thread: all training ran on academic compute (JHU DSAI support). The project page details the pipeline—depth-estimated colored point clouds, Blender-authored motion handles, per-frame background/foreground XYZ and identity maps encoded by a frozen Wan VAE, conditioned into the frozen diffusion model via a half-resolution side branch and LoRA adapters, with only those components trained.
Related event: GenCine: Single-Image Video Generation with 3D Motion Control(3 posts)→
More from Multimodal
- Seedance 2.5 one-shot chase video wows X, full prompt shared — umesh_ai · 2026-10-03
- A ComfyUI workflow for retexturing Makehuman characters in Blender keeps improving — o0ANARKY0o · 2026-10-03
- AI video praised for standout screenwriting and cinematography; ElevenLabs v4 nails Latin — ctjlewis · 2026-10-03
- Video-as-3D-object demo gets segmentation, letting you search any object in a scene — andy_matuschak · 2026-10-03
- LTX 2.5 Restore IC-LoRA restores archive footage to clean, colorized HD — TinfoilTricorn · 2026-10-03
- Photoshop's Remove Tool now erases shadows and reflections along with objects — jnack · 2026-10-03