DVD-AWQ — AWQ-quantized DVD video-depth DiT

AWQ (activation-aware weight-only) quantized versions of the DVD deterministic video-depth DiT (Wan2.1-1.3B backbone: 30 DiTBlocks, 300 linear layers). These are drop-in replacements for the full-precision DiT that shrink the checkpoint while preserving depth quality.

checkpoint	bits	size	vs bf16	drone AbsRel↓	drone δ1↑
(bf16 reference)	16	2.84 GB	1.0×	—	—
dvd_dit_awq_w8g128.safetensors	W8A16	1.49 GB	1.9× smaller	0.011	0.988
dvd_dit_awq_w4g128.safetensors	W4A16	0.79 GB	3.6× smaller	0.082	0.953

Metrics are computed against the bf16 model's own output on a held-out drone video (quantization fidelity), after per-video affine alignment. w8g128 is effectively lossless (masked δ1 = 1.000); w4g128 trades some accuracy for maximum compression, with the largest errors concentrated in near-zero-depth pixels.

Method

• Weight-only AWQ, group size 128, asymmetric (zero-point) group quantization.
• Activation-aware scaling on 4 mathematically-exact fusion groups per DiTBlock (norm3→cross_attn.q; self_attn.v→o; cross_attn.v→o; ffn GELU→ffn.2) plus weight clipping on 6 linears per block (q/k skipped).
• Calibration uses real block-0 activations captured from depth inference (the DiT runs single-pass at one fixed timestep with zeroed text context).
• w4g128 is calibrated on a diverse (DAVIS + navigation) set; w8g128 on navigation clips (it is calibration-insensitive).

Files

• dvd_dit_awq_w8g128.safetensors, dvd_dit_awq_w4g128.safetensors — packed low-bit DiT weights (nibble-packed int4 / int8 + per-group scales/zeros).
• awq_results_w8g128.pt, awq_results_w4g128_div.pt — the AWQ scale/clip search results, for re-packing or re-application.

Usage

Load with the packing/loader utilities in the code repository (quant_script/load_awq.py, --awq_ckpt flag on the inference script): https://github.com/yihao-liang/DVD (branch awq-quant-squashed). See quant_script/README.md there for the full reproduction and inference commands.

License & attribution

Code is Apache-2.0. These weights are a derivative of DVD and inherit its CC BY-NC 4.0 license — non-commercial, academic/research use only. Please cite:

@article{zhang2026dvd,
  title={DVD: Deterministic Video Depth Estimation with Generative Priors},
  author={Zhang, Hongfei and Chen, Harold Haodong and Liao, Chenfei and He, Jing and Zhang, Zixin and Li, Haodong and Liang, Yihao and others},
  journal={arXiv preprint arXiv:2603.12250},
  year={2026}
}

Quantization built on llm-awq (MIT).