comind-actoractor-wo-sharedenv (iter 12200)

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

What this experiment is

Ablation of comind_actoractor_personpose_shared_refpose_base2b that removes the shared environment context while keeping joint generation (V=2) and the two-person identity-coloured pose stream. Exactly two things change:

• Own-context-only warping -- control_input_warped / control_input_visibility are built from each view's own past pool. Frames the parent sourced from the partner's pool become holes (black frame, visibility all-False). Data: ego_gen_comind_single.
• No reference frames -- reference appearance, reference pose and posed-reference Plucker all off (R=0).

Consequence: the single-ego corpus stores cameras canonicalised to each view's own frame 0, so the shared canonicalisation is a no-op and Plucker is per-view; the pair no longer shares a world frame. Cross-view self-attention is the only remaining channel between the two egos.

Training

• Warm start: base Cosmos 2B multiview (nvidia/Cosmos-Predict2.5-2B).
• 1x NVIDIA H200 (fsdp_shard_size=1, grad_accum_iter=2, global batch 2, matching the 2-GPU base2b arm's global batch).
• Trained from scratch for this arm (no resume) to iter 12,200; checkpoints every 200 iterations. Training was stopped with SIGTERM shortly before the GPU allocation expired; this is the last saved checkpoint.
• Loss (rectified flow) is noisy, roughly 0.04-0.09 over the last iterations.

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 colleague's own copy of the cosmos-predict2.5 tree (predict2_multiview/configs/vid2vid/experiment/nymeria_pose_2actor.py, comind_actoractor_wo_sharedenv), which was not modified for this run.