- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
BTCache allocates a fixed computation budget jointly across transformer block groups and denoising steps. It combines residual reuse with damped extrapolation, a learned ensemble that predicts terminal latent, perceptual, and temporal risks, and an exact feasibility test that preserves the remaining budget and maximum cache age. Training uses complete continuation trajectories and policy-visited branches rather than only local feature errors. Experiments on HunyuanVideo, Wan2.1, and FLUX report roughly 3.4–3.6× end-to-end speedups. The idea is useful, but the evidence does not yet clearly isolate the benefit of joint block decisions or validate the terminal-risk predictor that drives them.
Critical strengths
- Explicit control of the computation budget. Proposition 1 and the feasible-action filter give a concrete way to enforce both the exact number of remaining FULL cells and cache-age limits. This is more useful for predictable resource allocation than a threshold that only controls average compute, and the paper appropriately reports wall-clock measurements separately.
- Supervision targets the eventual generation outcome. The same-prefix branching scheme and separate terminal-risk heads address error accumulation and tradeoffs that a local feature-distance proxy can miss. In Table 4, replacing terminal supervision with local-error supervision reduces HunyuanVideo PSNR from 17.23 to 16.54 dB, supporting this design choice.
- The implementation includes realistic overhead measurements. The main latencies include the controller, and Appendix D reports controller costs of 1.74 s on HunyuanVideo and 1.08 s on Wan2.1. Tests span image and video generation, and the matched-cost FLUX comparison improves PSNR over MagCache while preserving similar LPIPS. These results show that controller overhead does not automatically erase the savings.
Critical weaknesses
- The evaluation does not yet establish a robust quality advantage. Section 4.1 uses only nine HunyuanVideo prompts and eight Wan2.1 prompt–seed pairs. Several quality margins are small: HunyuanVideo VBench is 79.44 versus 79.43 for TaylorSeer, without paired uncertainty estimates. The seven-of-nine win count and budget sweep support the internal comparison but do not establish broader generalization. On 40 FLUX prompts, Table 3 reports 1.47 dB higher PSNR than matched-cost MagCache with identical LPIPS (0.3218) and nearly identical SSIM. The faster MagCache setting also has higher PSNR but worse LPIPS than its published setting. These patterns are not intrinsically inconsistent, but the missing FLUX operating parameters and per-sample results prevent checking their cause. Broader held-out evaluation and verification of this comparison are needed before drawing a strong quality-lead conclusion.
- The closest scheduling comparisons are missing. Section 2 explicitly identifies MeanCache and DPCache as the closest formulations, yet neither appears in the experiments. Table 4 and Appendices D–E do support finer allocation under the authors’ controller; they do not establish superiority over these independent fixed-budget schedulers. Figure 1d also provides evidence that one-step reuse errors are nonadditive, but it does not measure the benefit of learning joint interactions over an additive scorer with the same feasible partial-block masks. A budget or latency sweep against an applicable close scheduler, together with this controlled additive comparison, would test the claimed scheduling advance more directly. Any incompatibility should be explained and the comparative claim scoped accordingly.
- The data requirements and offline cost of the learned controller cannot be assessed. Section 3.5 describes the rollout procedure and Appendix C specifies the architecture, so training is not wholly undocumented. However, the paper does not report training prompt sources and counts, terminal rollout and branch counts, or total data-generation, fitting, and calibration cost per backbone. It specifies calibration/test separation without fully describing the training split’s relationship to those sets. These omissions matter for a five-member ensemble trained on terminal outcomes: they prevent assessing data dependence, generalization, and the number of generations needed to amortize preparation. The reported 2% online controller overhead does not resolve this cost question.
Questions for authors
- How does the joint-mask predictor compare with an additive per-group predictor using the same feasible masks, terminal supervision, residual operator, and budget? This comparison would separate the value of modeling group interactions from the benefit of allowing partial refreshes.
- On held-out policy-visited states, can you compare predicted action rankings with terminal outcomes obtained by branching and then following the deployed continuation? Please report the selected-action regret and explain whether another policy-evaluation or data-collection iteration changes the decisions.
- How does BTCache compare with a closely related fine-grained scheduler at the same computation budget and approximation policy? Please include the offline training and calibration cost when assessing whether the additional learned machinery is worthwhile.
2 comments
Response to Weakness 3
We agree that the data requirements and offline cost belong in the paper, and App. I now reports them.
| HunyuanVideo item | Value |
|---|---|
| Rollouts | Round 1: 1200(30 prompts × 10 families × 4), about 180 GPU-h. Round 2: 60, about 9 GPU-h; about 27 GPU-h together. A800 with CPU offload; about 8 min per rollout, 23 min per all-full reference |
| Predictor | 5 members, independent seeds; AdamW, lr 8e-4, weight decay 1e-4, clip 5.0; one family per batch with full recurrent backprop; 600 updates per member in round 1, 900 in round 2; CPU, about 35 min in round 1 |
| Loss | Gaussian NLL on standardized log terminal risk ×1.0, same-state pairwise rank ×0.5, local NLL ×0.25, next-state smooth-L1 ×0.25 |
| Comparator | 5 members, same optimizer |
We fixed 12 states visited by the selector on HunyuanVideo before seeing any result. At each state we forced each of the 16 masks , followed BTCache to the final step, and scored PSNR against the all-full video. Repeating the selector's own branch reproduces it exactly in 10 of 12 states. Across the 16 masks, PSNR spans 2.14 dB on average.
| State | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Best mask − best timestep-level mask (dB) | 0.23 | 0 | 0.09 | 0.20 | 0 | 0 | 0.19 | 0.18 | 0 | 0.40 | 0.12 | 0.05 |
| Partial masks better than best timestep-level (of 14) | 1 | 0 | 3 | 10 | 0 | 0 | 8 | 6 | 0 | 1 | 3 | 2 |
| Regret of the executed mask (dB) | 0 | 0 | 0.09 | 0 | 0 | 0 | 0.19 | 0.18 | 0 | 0. | 0.08 | 0.05 |
To measure how much a better predictor could gain, we took 25 states (the 12 above and 13 round-3 states) and compared the best of the 16 masks with the best timestep-level mask under the same exact-budget continuation.
| Metric (states) | Mean | Median | Partial mask best in |
|---|---|---|---|
| PSNR (25) | 0.20 dB (max 1.97) | 0.085 dB | 82% |
| LPIPS (25) | 0.1046 | 0.0004 | 80% |
| Temporal (13) | 0.193 | 0.061 | 75% |
We thank the reviewer for the thorough review and for recognizing the explicit budget control and the supervision on terminal outcomes. Because the questions were specific, we could run most of them as stated. These include the additive scorer under identical masks, ranking and regret on held-out policy-visited states, and a budget sweep against close schedulers with offline cost. Some of the results lead us to state our claims more narrowly, and each answer says where.
Response to Weakness 1
We agree that margins of this size need paired uncertainty estimates. As an additional evaluation, we extended HunyuanVideo to 36 test prompts at 12 NFE (48 of the 200 group–step cells; one timestep is 4 cells), one fixed seed per prompt. The prompts are the paper's 9 test prompts, 15 fixed VBench prompts and 12 further VBench prompts. We compare with MagCache configured to compute exactly 12 steps (MagCache-12; the stronger of two such settings, schedule {0–6, 13, 19, 28, 37, 46}). We evaluate BTCache; the reference plan π0 re-selected after a second round of data on the training prompts
As an additional evaluation, we ran MagCache with its official forward pass on the 40 DrawBench prompts, with one fixed seed per prompt and the same all-full references as BTCache. In these runs BTCache has higher PSNR and SSIM and lower LPIPS than MagCache at every setting below.
Response to Weakness 2 & Question 1
Thank you for suggesting this controlled comparison. We trained an additive per-group predictor, which is the joint predictor with the cross-group interaction removed, on identical training episodes and with the same recipe, reference plan, β, selector, budget and residual operator. The additive predictor gives lower PSNR and SSIM and higher LPIPS than the joint predictor.
Response to Weakness 2 & Question 3
we ran a budget sweep against DPCache and BudCache on FLUX.1-dev, compared with BWCache on HunyuanVideo, and now state our comparative claims more narrowly. BTCache has the best PSNR, SSIM and LPIPS.
BudCache's official search takes 26 min (8 NFE) and 43 min (14 NFE) on our hardware. ECAD's released schedules sample 20 steps against our 50-step reference. BWCache's official code covers Open-Sora, with Wan and HunyuanVideo listed as future work, so we re-implemented it. MeanCache's official code covers Qwen-Image and Z-Image, and ProCache's covers DiT, FLUX and PixArt; we could not run either in the time available. DPCache's official code covers FLUX and Wan2.1-I2V-14B-720P only, and porting it to HunyuanVideo and Wan2.1-T2V-1.3B with its PACT calibration could not be completed in the rebuttal period.