comind-actoractor-personpose-refpose-singlestream (iter 14400)

Checkpoint from the comind_actoractor_personpose_shared_refpose_singlestream experiment, continued as a guest run on a teammate's behalf (original experiment authored by a colleague; this run picks it up from their iter_000005400 checkpoint and continues training under a different account due to GPU availability).

What this experiment is

Ablation of comind_actoractor_personpose_shared_refpose: the two ego streams (leader/helper) are generated independently instead of jointly. One CoMind clip becomes TWO V=1 samples (leader, helper) rather than one V=2 sample. Per-stream conditioning is unchanged from the joint run (own vae_concat pose, own warped/visibility, own R=4 slice of the greedy SHARED reference set, matching reference_pose skeletons, and a Plücker ray map) -- the ablation isolates exactly two things that come from denoising the pair together:

• cross-view self-attention: a stream can no longer attend to the other ego's tokens (or reference frames).
• the shared canonical frame: with V=1 the Plücker map's "target view0, frame 0" is just the stream's own frame 0, so the two streams are no longer grounded in a common space.

Nothing is dropped from the per-stream inputs -- this is not a conditioning ablation, it isolates what joint generation buys over generating each ego stream independently.

Training

• Warm start: base Cosmos 2B multiview (nvidia/Cosmos-Predict2.5-2B).
• Continued from iter_000005400 (the colleague's own prior progress) through iter 14,400 in this run.
• 2x NVIDIA H200, fsdp_shard_size=2, checkpoints saved every 400 iterations.
• View embedding frozen (nothing for it to distinguish with V=1 streams).

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

This checkpoint continues a colleague's in-progress experiment as a guest run on borrowed GPU time; the training code lives in their own copy of the cosmos-predict2.5 tree (predict2_multiview/configs/vid2vid/experiment/nymeria_pose_2actor.py, comind_actoractor_personpose_shared_refpose_singlestream), which was not modified for this run.