cs.CVMar 13, 2026
SaveDraft-and-Target Sampling for Video Generation Policy
Organizations: South China Normal University · University of Western Australia
Abstract
Video generation models have been used as a robot policy to predict the future states of executing a task conditioned on task description and observation. Previous works ignore their high computational cost and long inference time. To address this challenge, we propose Draft-and-Target Sampling, a novel diffusion inference paradigm for video generation policy that is training-free and can improve inference efficiency. We introduce a self-play denoising approach by utilizing two complementary denoising trajectories in a single model, draft sampling takes large steps to generate a global trajectory in a fast manner and target sampling takes small steps to verify it. To further speedup generation, we introduce token chunking and progressive acceptance strategy to reduce redundant computation. Experiments on three benchmarks show that our method can achieve up to 2.1x speedup and improve the efficiency of current state-of-the-art methods with minimal compromise to the success rate. Our code is available.
Figures & tables
Figure 1: The Draft-and-Target Sampling Framework. In the first stage of DTS, the draft sampling generates a coarse denoising trajectory by taking large steps in a sequence manner, which provides a fast approximation of the denoising trajectory. Then the target sampling refines this trajectory in parallel by taking small steps, which leads to a more precise denoising trajectory. Finally, in verification stage we compare these two trajectories to determine if tokens are accepted or rejected. If a token is rejected, the draft sampling restarts from this position with its corresponding target token as the new starting point.
Figure 2: Token Chunking. In naive DTS, the draft sampling generates a long sequence in one pass, while token chunking divides the draft trajectory into smaller chunks that are processed chunk by chunk.
| Method | Kitchen | Living Room | Bedroom | Bathroom | Overall | Speedup | NFEs |
|---|---|---|---|---|---|---|---|
| BC-Scratch* | 1.7 | 3.3 | 1.7 | 1.7 | 2.1 | / | / |
| BC-R3M* | 0.0 | 0.0 | 1.7 | 0.0 | 0.4 | / | / |
| AVDC-100* | 26.7 | 23.3 | 38.3 | 36.7 | 31.3 | 1x | 84800 |
| AVDC-100 | 30 | 16.7 | 35 | 26.7 | 27.1 | 1x | 84800 |
| AVDC-50 | 21.7 | 21.7 | 25 | 20 | 22.1 | 2.0154x | 45950 |
| AVDC-25 | 26.7 | 21.7 | 23.3 | 18.3 | 22.5 | 3.9593x | 22525 |
Table 1: Comparison results on 4 types of rooms in iThor. We reproduce AVDC with 100 denoising steps as in the original paper as our baseline, while AVDC-10 is a faster variant that uses only 10 denoising steps. * denotes the reported results from the original paper.
Figure 3: Meta-World: faucet-close. In this task, the agent is required to rotate the faucet handle and close it. As shown in comparison (d), the trajectory generated by DDIM-10 is markedly different from DDIM-100, with the faucet being rotated in an entirely different direction. As shown in comparison (e), our trajectory closely align with DDIM-100, both successfully closing the faucet.
| Task | BC-Scratch* | BC-R3M* | AVDC* | AVDC | AVDC-10 | + Ours |
| door-open | 21.3 | 1.3 | 72 | 72 | 69.3 | 69.3 |
| door-close | 36 | 58.7 | 89.3 | 97.3 | 100 | 97.3 |
| basketball | 0.0 | 0.0 | 37.3 | 33.3 | 16 | 26.6 |
| shelf-place | 0.0 | 0.0 | 18.7 | 21.3 | 5.3 | 4 |
| btn-press | 34.7 | 36 | 60 | 52 | 52 | 45.3 |
| btn-press-top | 12 | 4 | 24 | 14.7 | 4 | 18.6 |
Table 2: Comparison results on 11 tasks in Meta-World. We reproduce AVDC with 100 denoising steps as in the original paper as our baseline, while AVDC-10 is a faster variant that uses only 10 denoising steps. Following the original setting, we evaluate each task for 25 trials with 3 camera poses, 825 in total. * denotes the reported results from the original paper.
| Task | GCP* | GCP-P | GCP-I | DDIM10 | DPM | +Ours |
| put-red-mug-left | 38.4±15.3 | 45 | 30 | 40 | 30 | 30 |
| put-red-mug-right | 40.8±7.8 | 40 | 40 | 25 | 40 | 40 |
| put-white-mug-left | 51.2±3.9 | 55 | 40 | 50 | 45 | 70 |
| put-Y/W-mug-right | 38.4±8.6 | 30 | 40 | 40 | 35 | 25 |
| put-choc-left | 70.4±12.8 | 70 | 50 | 45 | 50 | 65 |
| put-choc-right | 79.2±3.9 | 80 | 70 | 85 | 80 | 90 |
Table 3: Comparison results with different solvers on 8 tasks across two different scenes in Libero. GCP* denotes the baseline method goal-conditioned policy and the reported results are directly taken from the original paper. GCP-P and GCP-I represent reproduced results using DDPM and DDIM samplers with 100 denosing steps, respectively. DDIM10 represents DDIM with 10 denosing steps and DPM represents the DPM-Solver with 10 denosing steps. The success rates are in increments of 5% because each task is evaluated over 20 trials, leading to discrete averages.
| Task | Chunk-2 | Chunk-4 | Chunk-6 | Chunk-8 | Chunk-10 |
|---|---|---|---|---|---|
| put-red-mug-left | 25 | 25 | 40 | 30 | 35 |
| put-red-mug-right | 45 | 45 | 50 | 55 | 40 |
| put-white-mug-left | 40 | 60 | 60 | 45 | 50 |
| put-Y/W-mug-right | 25 | 25 | 30 | 30 | 15 |
| Overall | 33.75 | 38.75 | 45 | 40 | 35 |
| Time cost(s) | 5.5872 | 4.9862 | 5.4128 | 5.6836 | 6.8004 |
Table 4: Ablation on 8 tasks across two different scenes in Libero. We evaulate different chunk length.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | Kitchen | Living Room | Bedroom | Bathroom | Overall | Speedup |
|---|---|---|---|---|---|---|
| Chunk-2 | 30 | 20 | 31.7 | 18.3 | 25 | 1.934x |
| Chunk-4 | 28.3 | 16.7 | 31.7 | 28.3 | 26.25 | 2.1107x |
| Chunk-6 | 26.7 | 23.3 | 38.3 | 36.7 | 31.3 | 2.1445x |
| Chunk-8 | 28.3 | 15 | 26.7 | 21.7 | 22.9 | 2.0655x |
| Chunk-10 | 16.7 | 30 | 33.3 | 31.7 | 27.9 | 1.9935x |
Table 5: Ablation on 4 types of rooms in iThor. We evaulate different chunk length.
| Task | Chunk-2 | Chunk-4 | Chunk-6 | Chunk-8 | Chunk-10 |
|---|---|---|---|---|---|
| door-open | 60 | 68 | 70.7 | 64 | 69.3 |
| door-close | 97.3 | 98.7 | 97.3 | 100 | 97.3 |
| basketball | 14.7 | 13.3 | 21.3 | 22.7 | 26.6 |
| shelf-place | 1.3 | 1.3 | 1.3 | 0 | 4 |
| btn-press | 53 | 60 | 50.7 | 42.7 | 45.3 |
| btn-press-top | 2.7 | 8 | 5.3 | 0 | 18.6 |
Table 6: Ablation on 11 tasks in Meta-World. We evaluate different chunk lengths.
| Method | Kitchen | Living Room | Bedroom | Bathroom | Overall | Speedup |
|---|---|---|---|---|---|---|
| Thre-4 | 21.7 | 23.3 | 15 | 23.3 | 20.8 | 1.8074x |
| Thre-6 | 31.7 | 20 | 38.3 | 23.3 | 28.3 | 1.8831x |
| Thre-8 | 30 | 20 | 36.7 | 21.7 | 27.1 | 1.8962x |
| Thre-10 | 30 | 23.3 | 33.3 | 16.7 | 25.8 | 1.8691x |
| Thre-12 | 25 | 23.3 | 33.3 | 20 | 25.4 | 1.889x |
Table 7: Ablation on 4 types of rooms in iThor. We evaulate different matching threshold.
| Task | Thre-4 | Thre-6 | Thre-8 | Thre-10 | Thre-12 |
|---|---|---|---|---|---|
| put-red-mug-left | 45 | 35 | 30 | 30 | 40 |
| put-red-mug-right | 45 | 40 | 40 | 40 | 50 |
| put-white-mug-left | 55 | 35 | 50 | 70 | 60 |
| put-Y/W-mug-right | 20 | 20 | 20 | 25 | 30 |
| put-choc-left | 45 | 70 | 50 | 65 | 70 |
| put-choc-right | 60 | 75 | 60 | 90 | 70 |
Table 8: Ablation on 8 tasks across two different scenes in Libero. We evaulate different matching threshold.
Figure 4: Meta-World: Basketball. In this task, the agent needs to place the basketball into the basket. As shown in comparison (d), although the trajectory generated by DDIM-10 is close to DDIM-100, a ghosting artifact appears in the second-to-last frame where an extra fragment of the robotic arm is generated near the true arm, and in the final frame the basketball is positioned slightly off-center relative to the basket, which leads to fail in this task. In contrast, our trajectory closely aligned with DDIM-100, both successfully placing the basketball directly above the center of the basket.
Figure 5: Meta-World: Button-press-topdown. In this task, the agent is required to press the button from a top-down direction. As shown in comparison (d), although the trajectory generated by DDIM-10 is close to DDIM-100, a ghosting artifact appears in the second-to-last frame where an extra fragment of the robotic arm is generated near the true arm, and in the final framethe button not completely pushed down. In contrast, our trajectory closely aligned with DDIM-100, both successfully press the button all the way down.
Figure 6: Meta-World: Hammer. In this task, the agent is required to use a hammer to drive a nail into the wooden block. As shown in the figure, the trajectory generated by DDIM-10 is different from DDIM-100, losing track of it in the final two frames and failing to drive the nail into the block. In contrast, our trajectory closely match DDIM-100, with both successfully embedding the nail into the block.
Figure 7: iTHOR: desk-lamp. In this task, the agent starts in a bedroom and must find the desk lamp among other furniture. As shown in the figure, both DDIM-10 and DDIM-100 fail to locate the desk lamp. In contrast, our method successfully finds the desk lamp and navigates toward it beyond the final frame.
More in this topic
Diffusion Model Inference Acceleration
Diffusion model inference acceleration reduces the time and compute needed to generate or reconstruct outputs while preserving quality.
See all 506 papers