Tevior/kvcontrol-c-concat

KV-Control (C-Concat substrate) — pretrained weights

Pretrained weights for KV-Control, a parameter-efficient attention-side key/value injection adapter for trajectory-controlled text-to-motion generation on a frozen PartVQ + C-Concat masked-motion (MaskGIT-style) backbone.

Code: https://github.com/CHDTevior/kvcontrol-c-concat

Download the whole tree into ./pretrained/ of the code repo (bash scripts/download_weights.sh).

Files

Path	Description
partvq/vq_net_best_fid.pth + partvq/skeleton_partition.json	Frozen part-aware VQ tokenizer + skeleton partition (required substrate)
normalization/mean.npy, normalization/std.npy	263-dim HumanML3D feature statistics
c_concat_base/net_best_fid.tar	Frozen C-Concat masked-motion backbone
kv_control/net_best_fid.tar	KV-Control trajectory adapter (best FID)
kv_control/net_best_top3.tar	KV-Control trajectory adapter (best Top-3)

Results (HumanML3D test, single-joint pelvis trajectory, 5-repeat mean ± 95% CI)

Protocol	FID	Top-3	KPS (cm)
M2 (MaskControl-matched TTT)	0.093 ± 0.011	0.795 ± 0.011	1.10 ± 0.01
M3 (+ dynamic refinement)	0.096 ± 0.011	0.793 ± 0.006	0.96 ± 0.01

Citation

Anonymous AAAI-2027 submission under review; author information withheld.

License

MIT (weights and code). Please also honor the licenses of HumanML3D and upstream projects (MoMask / MaskControl / OpenAI CLIP).