comind-actoractor-personpose-refpose-base2b (iter 5200)

Checkpoint from the comind_actoractor_personpose_shared_refpose experiment (base2b arm), continued as a guest run on a teammate's behalf (original experiment authored by a colleague; this run picks it up from their iter_000002800 checkpoint and continues training under a different account due to GPU availability).

What this experiment is

Joint (V=2) CoMind actor-actor generation: leader and helper ego streams are denoised together in one forward pass -- full cross-view self-attention (each stream attends to the other's tokens and reference frames) and a shared canonical camera frame (Plücker rays expressed in "target view0, frame 0"). Conditioning per stream: vae_concat identity-colored pose, own warped/visibility, an R=4 slice of a greedy SHARED reference set (with matching reference_pose skeletons), plus the cross-view Plücker ray map.

This is the base2b arm: warm-started directly from base Cosmos 2B (nvidia/Cosmos-Predict2.5-2B), as opposed to the nymwarm/singlestream arms which warm-start from a Nymeria-pretrained checkpoint instead. It is the joint-generation counterpart to kimhosoo/comind-actoractor-personpose-refpose-singlestream (same conditioning, but the two ego streams generated independently, V=1) -- meant for a joint-vs-independent comparison.

Training

• Warm start: base Cosmos 2B multiview (nvidia/Cosmos-Predict2.5-2B), per the base2b arm definition.
• Continued from iter_000002800 (the colleague's own prior progress) through iter 5,200 in this run.
• 2x NVIDIA H200, fsdp_shard_size=2, checkpoints saved every 400 iterations.
• Run stopped automatically (SIGTERM via timeout) ~20 minutes before the borrowed GPU allocation expired, to guarantee a clean final checkpoint.

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), which was not modified for this run.