- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
BASS constructs an offline mixed-precision policy for video diffusion models with auxiliary conditioning branches. It combines selective smoothing, sensitivity measured at enclosing block outputs, and a greedy allocation of W4A4 and W8A8 execution across structural groups and timestep contexts. Experiments on two Wan models report improved fidelity to BF16 generations and measured memory and latency savings.
Critical strengths
- Treating the conditioning branches as part of precision allocation is a useful problem to study. It is sensible to assess adapters together with the backbone, since their importance to control need not follow their parameter count.
- The reported improvements in fidelity to BF16 are substantial on Wan-Animate. The TikTok and Fashion UBC results agree across several paired metrics, and the visual examples support preservation of the reference appearance better than the reported baseline configurations.
- The smoothing and random-selection ablations are informative. Selective smoothing outperforms applying smoothing everywhere, while sensitivity-based allocation improves over random selection at the stated budget. These results give useful initial support for the complete recipe.
- The Wan-VACE results in the appendix add a useful second conditioning architecture. The paper also describes an actual GPU execution path and measures end-to-end latency, which helps assess whether the quantization recipe can be used in practice.
Critical weaknesses
- The main comparisons do not establish equivalent compression costs. Uniform W8A8 is also slightly faster than BASS in Figure 4, but its quality is not reported. I would like to see whether BASS offers a better quality and memory tradeoff than this simple alternative and comparably costly mixed-precision baselines.
- The allocation cost is defined as byte savings, but its relation to resident memory is unclear when one layer uses different transformed weights across timestep contexts. The reported memory reduction needs accounting for cached weight versions and FP8 expansion buffers. Otherwise the optimized budget and measured deployment benefit remain disconnected.
- The ablations support selective smoothing and sensitivity-based allocation, but leave the central structural choices less clear. In particular, keeping adapters uniformly at higher precision and measuring sensitivity at layer outputs are missing controls. These would help show whether the gains come from the proposed branch-aware block assessment.
- The evaluation primarily measures agreement with BF16 generations and general video quality. It does not directly test reference-identity preservation or adherence to the driving pose, despite these motivating the branch-conditioned setting. Targeted identity and motion-control measurements are needed to establish that the most important conditioning behavior survives compression.
Questions for authors
- Which quality results correspond to the ten-step configuration used for timing? The appendix describes twenty steps, so please clarify how the allocation is mapped or recalibrated for the reported deployment workload.
- How many clips enter the evaluation, and were calibration and budget selection separated from those clips? The small VACE differences are especially difficult to interpret without these counts and the corresponding variability.
- What is the offline profiling cost relative to the evaluated baselines? Measuring groups separately across timestep contexts could make policy construction expensive, even if the resulting policy is fixed during inference.
- How does the block-output sensitivity measure differ from that in 6Bit-Diffusion? The application to control branches is interesting, and a clearer account of what this assessment adds would sharpen the contribution.
4 comments
W4: Targeted identity and motion-control measurements are needed to establish that the most important conditioning behavior survives compression.
A4: Thank you for the suggestion on conducting experiments to test the preservation of identity and motion. We add experiments for BF16, uniform W8A8, BASS, and the three adapted baselines on TikTok dataset. The TikTok clips contain severe body movements, making them a challenging test of motion adherence and identity detectability in generated videos. We evaluate identity against the input reference image and pose against the frame-aligned driving control. The mixed-precision baselines (ViDiT-Q, Q-VDiT) adopt the same ratio of mixed W4A4/W8A8 quantization as BASS.
CSIM is the cosine similarity of ArcFace identity embeddings on detected faces. AKD is the mean Euclidean error of jointly detected keypoints in pixels, reported separately for the body and hands. MKR is the ratio of valid driving keypoints missing from the generated video. Face coverage reports the ratio of generated frames with a detected face.
Table 3. Identity and motion preservation on TikTok.
| Method | CSIM ↑ | Body AKD ↓ (px) | Hand AKD ↓ (px) | Body MKR ↓ (%) | Hand MKR ↓ (%) | Face coverage ↑ (%) |
|---|---|---|---|---|---|---|
| BF16 | 0.4889 | 7.88 | 17.66 | 1.37 | 11.95 | 99.48 |
| W8A8 | 0.4622 | 7.91 | 14.06 | 3.24 | 11.86 | 99.22 |
| ViDiT-Q | 0.4659 | 7.98 | 14.15 | 3.88 | 13.18 | 99.48 |
| Q-VDiT | 0.4663 | 7.91 | 13.74 | 2.98 | 11.99 | 99.22 |
| SmoothQuant | 0.0119 | 30.57 | 42.55 | 41.98 | 59.18 | 10.91 |
| BASS | 0.4664 | 7.71 | 15.08 | 2.17 | 11.12 | 99.48 |
Among all baselines, BASS performs better in the similarity of identity and detected body keypoints while falls a little behind in detected hand keypoints. What's more, BASS's identity similarity as well as body and hand error remain closely to BF16, demonstrating the preservation of motion and identity.
Q1: Which quality results correspond to the ten-step configuration used for timing? The appendix describes twenty steps, so please clarify how the allocation is mapped or recalibrated for the reported deployment workload.
A5: We apologize for the mismatch of our experiment configurations and the configuration used for timing. All of the proposed quality results are conducted with a 20-step inference procedure, which is the default setting of Wan2.2-Animate. The timing experiment is conducted under 20-step configuration and stops at step 10 for a quick feedback. Since the inference time cost for each step remains stable, it can also reflect the speed of proposed methods. We report the actual allocation of the policy used in our supplemental experiment at target cost ratio 0.85. The quantizable scope contains 480 backbone Linear layers and 24 adapter Linear layers.
What's more, we report the W4A4 ratio of actual layer calls in the whole 20-step inference.
| Component | W4A4 ratio of actual layer calls |
|---|---|
| DiT backbone | 3010/9600 (31.35%) |
| Adapter | 90/480 (18.75%) |
| Overall | 3100/10080 (30.75%) |
Q2: How many clips enter the evaluation, and were calibration and budget selection separated from those clips?
A6: For each dataset, we choose 100 clips to enter the evaluation. The 5 clips used for calibration and budget selection is separated from the evaluation clips. We think the number of the clips we use for evaluation is large enough to avoid variability in results.
Q3: What is the offline profiling cost relative to the evaluated baselines?
A7: We thank you for the concerns on the offline profiling cost. Among the baselines, the cost of the offline profiling procedure of BASS is similar to ViDiT-Q which also conducts an assessment of sensitivity, larger than SmoothQuant and Q-VDiT. Though the offline profiling stage costs high, it is conducted once or few times for practical video generation.
Q4: How does the block-output sensitivity measure differ from that in 6Bit-Diffusion?
A8: Thank you for pointing out this connection. First, 6Bit-Diffusion defines sensitivity by quantizing one linear layer at a time, while our BASS evaluates structural groups, including simultaneous Q/K quantization, at the enclosing block output. This directly measures the combined perturbation of coupled projections through attention and subsequent block operations.
Second, 6Bit-Diffusion fits a predictor using the preceding timestep’s block input–output difference to route activation precision online. BASS aggregates calibration errors and then combines these risk estimates with cost savings to construct an offline policy under a shared backbone–adapter budget. This provides a measured basis for deciding where low precision is acceptable across conditioning and backbone pathways, without uniformly protecting or compressing all adapters. Inference executes the resulting group–context assignments through policy lookup and no extra calculation.
Thank you for the response and the additional comparisons. The W8A8 results clarify the quality trade-off, and the identity and pose results address my concern about conditioning. The fixed-adapter comparison supports joint precision allocation across the backbone and adapters. The benefit of measuring sensitivity at the block output rather than at individual layers remains unclear. I will maintain my score.
We are encouraged by your acknowledgement of our additional comparisons, and appreciate you for pointing out the missing illustration. We would like to clarify the advantages of adopting block-output sensitivity over layer-output sensitivity.
The distinction is between measuring the quantization error introduced at a layer and measuring its effect after propagation through the remaining operations within the block. Let h denote an intermediate layer output, δh its quantization error, and F the subsequent computation within the block. Layer-output error measures ‖δh‖², whereas block-output error measures ‖F(h + δh) − F(h)‖². Under a first-order approximation, the latter becomes ‖Jδh‖², where J is the Jacobian of F evaluated at h. This captures how subsequent computations can increase or reduce the effect of quantization errors on the block output. Consequently, similar layer-output errors can have different effects on the representation passed to the next block.
This distinction is particularly relevant to attention: errors in Q/K projections affect attention logits, but their effect on the block output also depends on softmax, value aggregation, and subsequent projections. Measuring at the enclosing block output evaluates the error after these operations. We therefore adopt block-output error in BASS as a computationally tractable sensitivity proxy.
We thank the reviewer for the thoughtful and constructive feedback. Your comments have guided our supplementary analyses and revisions. We provide detailed responses to each weakness and question below.
W1: The main comparisons do not establish equivalent compression costs. Uniform W8A8 is also slightly faster than BASS in Figure 4, but its quality is not reported.
A1: We thank you for the catch on the missing report of W8A8. We report the performance of W8A8 and compare it with our BASS as below:
Table 1. Uniform W8A8 and BASS on TikTok and Fashion UBC.
On both datasets, BASS outperforms W8A8 mainly in Subject Consistency, Background Consistency and Temporal Flickering, with similar results in Aesthetic Quality and Imaging Quality. For frame-level agreement metrics, BASS performs better than W8A8. We also conduct experiments and show the BASS's advantage over W8A8 in motion and identity detection metrics, you can check it in A4 Table 3.
What's more, we appreciate you point out the missing compression cost clarification in our paper. For mixed-precision baselines in our paper, we set them with the same actual W4A4/W8A8 ratio as BASS, so the results in Table 1 in our paper are the quality results of comparably costly mixed-precision baselines. We add this clarification in our revised Appendix A.3.
W2: The allocation cost is defined as byte savings, but its relation to resident memory is unclear when one layer uses different transformed weights across timestep contexts.
A2: Thank you for pointing out this distinction. Our allocation cost measures the quantized-weight bit-width cost, normalized to uniform W8A8. It provides a common precision budget for comparing allocation policies. We agree that it is not an reflection of the real resident or peak GPU memory.
Our implementation manages GPU memory through CPU offloading, stage-wise unloading of inactive modules, and on-demand transfer of quantized weights for computation. The low-precision transform is shared across timestep contexts for each physical layer, avoiding a separate transformed weight copy for every timestep.
W3: In particular, keeping adapters uniformly at higher precision and measuring sensitivity at layer outputs are missing controls. These would help show whether the gains come from the proposed branch-aware block assessment.
A3: We conduct a comparison experiment as you suggest. For the comparator, we fix the adapter at W8A8 and only conduct BASS's quantization mechanism on the backbone. The comparator shares a same overall ratio of W4A4/W8A8 with BASS. The results are shown below:
Table 2. Fixed-adapter versus BASS on TikTok and Fashion UBC.
Compared with the fixed-adapter policy, BASS shows better performance of VBench Average on both datasets. Moreover, BASS improves LPIPS and PSNR, while reducing FVD. This comparison provides evidence for the effectiveness of branch-aware assessment mechanism.