VPP2 object-video model, RoboTwin clean, step 10000
Private disaster-recovery backup of the completed object-video Video DiT. Input: nine full-scene RGB condition frames plus a complete action plan. Output: sixteen object-only future frames on a black background. Each frame composes observer left and robot ego right, 448 x 224. The model was initialized from VPP2 Stage-2 and trained for 10,000 optimizer steps.
step_010000.pt is a full BF16 Video DiT checkpoint, not a LoRA and not a full
optimizer-state resume. See checkpoint.json for the SHA-256 and pinned sources,
and RECOVERY.md for the recovery scope and omitted dependencies.
Project code: https://github.com/elephantsbug/vpp2-obj-generation
The VAE, CLIP, text encoder, videos, and text embedding tensors are not included. This is a private backup, with no new license grant; upstream terms still apply.