Modelpublic

dreamdifferent/vam-cross-level4-kuka-iiwa14-widowx-texture-v2-video-lora-iter200

sourceHugging Faceupdated 25d agoView on Hugging Face

0likes11downloads

Model Card

VAM-Cross two-camera MimicVideo Video2World LoRA

This public repository contains the trainable-only Video2World LoRA checkpoint from iteration 200 of v2w_kuka_iiwa14_level4_widowx_texture_2cam_hstack_v2_from_widowx250_video_fused_f0cea76_lora_r256. The source run has termination status completed; the selected four-component checkpoint set was verified before selecting the model weight.

Required initial backbone

—Kind: fused_video2world_dit
—Repository: dreamdifferent/widowx250-video-fused
—Revision: f0cea76b62c5dd66b06b9f965932ddea32a7b546
—Path: checkpoints/video_backbone/iter_000001060_fused.pt
—Bytes: 3913057284
—SHA-256: d0f24c049bee63b03d3b62747b240a2d1822ddd5f83a52fcd866a882e80122b1
—Source iteration: 1060

This is an adapter checkpoint, not a standalone fused model. Load the exact initial Video2World backbone above first and then apply this LoRA checkpoint. For fused_video2world_dit, the initial checkpoint already includes the earlier WidowX/Bridge LoRA fusion; loading the original Bridge backbone instead would be incorrect.

Supporting runtime artifacts

—MimicVideo commit: e3355dbc93132b576c02f920a59b4fc18a4f5906
—Checkpoint bundle: jonpai/mimic-video@f28339034831e3c2374be075e622e1ff38ebe0f8
—Video tokenizer: video_backbone/tokenizer/tokenizer.pth
—T5 text encoder directory: text_encoder/t5-11b

Use the same MimicVideo code and config recorded in config.yaml, vam_cross_video2world_config.json, and vam_cross_video_lora_manifest.json.

Training data contract

—Dataset revision: dreamdifferent/vam-cross-level4-kuka-iiwa14-widowx-texture@008d94dbe6fae4e6b8599cabf4515bd305f2d202
—Episodes/frames: 192 / 54749
—Cameras: observation.images.corner_cam, observation.images.front_cam
—View layout: hstack at 5 Hz
—Tasks: 29 episode-conditioned instructions (listed in vam_cross_video_lora_manifest.json)

The dataset itself is not included. Users must comply with the dataset's current access policy and the upstream MimicVideo, NVIDIA Cosmos, and base-checkpoint terms.