VAM-Cross MimicVideo fused Video2World DiT
This public repository contains a fused Video2World DiT produced by applying
the trainable-only LoRA from iteration 1060 of
v2w_widowx_target_2cam_wrist_ablation_v1_bridgeinit_lora_r256. The fused checkpoint is
checkpoints/video_backbone/iter_000001060_fused.pt.
The initial backbone and LoRA adapter are no longer separate at load time. This is a fused DiT checkpoint, not a complete inference bundle: the compatible MimicVideo code, video tokenizer, and T5 encoder are still required.
Provenance
- MimicVideo commit:
e3355dbc93132b576c02f920a59b4fc18a4f5906 - Initial backbone kind:
bridge_finetuned_fused - Initial backbone repository:
jonpai/mimic-video - Initial backbone revision:
f28339034831e3c2374be075e622e1ff38ebe0f8 - Initial backbone path:
video_backbone/v2w_bridge_lora_rank256_lr1.778e-04_bsz64_iter_000070043_fused.pt - Initial backbone SHA-256:
2723645edb158661af6d8e1c623ec21a63626d96e959799abb0b758397ac8c88 - Source dataset revision:
dreamdifferent/vam-cross-target-widowx250-native-corner-wrist@577a7e8b083dbdc267aec4eb84c109730884e1ce - Training tasks: 1
- Cameras:
observation.images.corner_cam, observation.images.wrist_cam - Video rate: 5 Hz
The private training dataset is not included. Users must comply with the upstream MimicVideo, NVIDIA Cosmos, and base-checkpoint terms.