comind-jointgen-noshared (iter 6500)

Checkpoint from the comind_jointgen_noshared experiment, run as a guest run on a teammate's behalf (experiment authored by colleagues; trained under a different account on borrowed GPU time).

What this experiment is

Ablation of comind_actoractor_personpose_shared_refpose: every input is single-ego, while generation stays joint (the two ego streams are still one V=2 sample, so cross-view self-attention is intact). It completes a 2x2:

conditioning	generation	experiment
joint	joint	comind_actoractor_personpose_shared_refpose
joint	independent	comind_actoractor_personpose_shared_refpose_singlestream
single-ego	joint	comind_jointgen_noshared (this checkpoint)

Cross-view conditioning removed (via the ego_gen_comind_single export):

• no reference frames (no shared appearance set, posed-reference Plucker, or reference_pose skeletons);
• warped condition from own past only (partner-sourced pixels become holes);
• self-only pose condition (own hands only, identity palette kept);
• per-view Plucker (cameras canonicalised to each view's own frame 0, so the views share no common world frame).

The reference modules are off in the net, so the model has fewer parameters than the parent and its raw loss is not comparable to the parent's.

Training

• Warm start: base Cosmos 2B multiview (nvidia/Cosmos-Predict2.5-2B).
• 1x NVIDIA H200 (fsdp_shard_size=1, grad_accum_iter=4, global batch 4 as in the parent).
• Trained to iter 6,500 (checkpoints every 500 iterations). The run stopped automatically shortly before the borrowed GPU allocation expired; this is the last saved checkpoint. The periodic validation-preview callback was disabled for this run (a dataset attribute mismatch, .clips vs .pairs, crashed it); this does not affect the weights.
• Data: dhyun22/ego_gen_comind_single.

Files

• model_ema_bf16.pt -- EMA weights in bfloat16 (recommended for inference), converted from the training-time DCP/FSDP checkpoint via scripts/convert_distcp_to_pt.py in the cosmos-predict2.5 repo.

Note

The training code lives in the colleagues' own copy of the cosmos-predict2.5 tree (predict2_multiview/configs/vid2vid/experiment/comind_jointgen_noshared.py), which was not modified for this run.