- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes a post-training quantization framework for video diffusion transformers, noting that quantization errors accumulate continuously during the denoising process, making it difficult to predict the final generation quality based on local quantization errors. PulseQuant introduces isolated block–timestep pulses during the diffusion process. The propagation gains estimated through quantization-error propagation are incorporated as propagation-risk weights in row-radius calibration. Second-order moments of the activations are estimated using calibration data, and two directions with high activation energy are identified through an iterative process. Neighboring-code edits are then performed along these two activation directions to reduce response error. On the Wan models, Self Forcing, and MiniMax-H3, PulseQuant improves dense reference fidelity metrics (PSNR/SSIM/LPIPS) and overall consistency at the W4A4 level, remains competitive at the W4A6 level, and demonstrates the advantages of native NVFP4 deployment as well as complementary acceleration with low-bit attention.
Critical strengths
-
Using isolated block–timestep pulses to measure the propagation of quantization error across multiple time steps is an effective method for determining the selection of weight radii. This approach addresses the issue that optimizing for local quantization error does not effectively translate to optimizing for the final quantization error. The two selected directions with high activation energy also effectively guide the editing of the neighboring-code.
-
Evaluation results across multiple video generators demonstrate that PulseQuant exhibits strong quantitative performance compared to baselines such as OrbitQuant and DVD-Quant. Ablation experiments also validate the effectiveness of each module. Deployment results on an RTX 5080 with native NVFP4 and two attention backends further show reductions in peak memory usage and latency.
Critical weaknesses
-
When performing Response-Subspace Correction, selecting only two directions with high activation energy may lead to insufficient calibration when activation energy is relatively dispersed. Neighboring-code edits also allow adjustments only within adjacent ranges, which may prevent the calibration from achieving its full potential.
-
Complex calibration processes—such as the injection of quantized pulses at specified block–timesteps, row-radius searches, the selection of high-energy activation directions, and neighboring-code edits—introduce calibration time overhead. This overhead may increase significantly, particularly as the number of anchor grids and the number of full-precision BF16 denoising steps increase.
Questions for authors
-
If a larger range of index shifts were allowed, or if more active directions were protected, would the final quantization quality improve further? Would this improvement justify the additional calibration cost?
-
What is the offline calibration cost? Can the authors analyze the calibration time for at least one model? How does the calibration cost vary with the anchor grid and the number of full-precision denoising steps ?
-
How sensitive are the results to the choice of the OrbitQuant codebook? Is subspace correction still effective when using a standard uniform quantizer per channel or other codebooks?
3 comments
Response to W2&Q2: Offline calibration cost.
We measured Wan-2.1-T2V-1.3B W4A6 calibration on a single NVIDIA A800 80GB GPU with three calibration prompts. With a anchor grid, denoising steps, and propagation horizon , calibration computation takes 713.41 s (0.198 GPU-hours): 603.53 s for risk profiling, 89.11 s for activation collection, 5.03 s for radius selection, 12.05 s for subspace correction, and 3.68 s for other operations. Including model loading, warm-up, cache writing, and process startup, the total is 747.32 s (0.208 GPU-hours).
The cost scales as follows (total GPU-hours, other settings at their defaults):
| Varied setting | Values | Total GPU-hours |
|---|---|---|
| Anchor grid (, ) | , , | 0.172, 0.208, 0.280 |
| Denoising steps (, ) | 10, 25, 50 | 0.171, 0.185, 0.208 |
| Horizon (, ) | 1, 2, 4, 8 | 0.176, 0.188, 0.208, 0.249 |
Profiling dominates the cost and grows sublinearly with the anchor grid and with in this range, while radius selection and correction are comparatively inexpensive. We will add these measurements to the Appendix.
Response to Q3: Subspace correction with a uniform quantizer.
We tested this directly on Wan-1.3B by replacing the OrbitQuant weight codebook with a per-output-channel uniform INT4 quantizer, while keeping the rotation, risk-weighted scale selection, and the A6 activation quantizer.
| Uniform INT4 | IQ ↑ | AQ ↑ | MS ↑ | DD ↑ | BC ↑ | SC ↑ | PSNR ↑ | SSIM ↑ | LPIPS ↓ |
|---|---|---|---|---|---|---|---|---|---|
| Without correction | 61.43 | 61.94 | 98.24 | 59.72 | 94.95 | 93.39 | 15.05 | 0.506 | 0.479 |
| With correction | 62.28 | 62.31 | 98.24 | 62.50 | 95.09 | 93.67 | 15.08 | 0.509 | 0.472 |
Correction improves or matches all nine metrics: PSNR rises from 15.05 to 15.08, SSIM from 0.506 to 0.509, and LPIPS improves from 0.479 to 0.472, with larger gains on IQ (+0.85) and DD (+2.78). Subspace correction therefore remains effective beyond the OrbitQuant codebook, under a standard per-channel uniform quantizer.
Response to W1&Q1: Additional correction directions and wider code-index edits.
We evaluated both extensions on Wan-1.3B W4A6: more correction directions (k=1,2,4, with the edit range fixed at ±1) and wider code-index edits (±1, ±2, ±4, with k=2). All runs share the same calibration prompts, risk profile, radius-selection procedure, and 326 evaluation prompts with matched seeds and BF16 references.
More directions. The differences are small and metric-dependent. Relative to k=2, k=4 improves four of nine metrics (AQ, MS, SC, and PSNR) and slightly worsens the other five (IQ, BC, DD, SSIM, and LPIPS); k=1 has the best LPIPS. k=2 is best on IQ, DD, BC, and SSIM, so we keep it as a balanced default rather than claim it is optimal for every metric.
Wider edits. Expanding from ±1 to ±2 improves AQ, BC, SC, PSNR, and SSIM, but worsens IQ, DD, and LPIPS; expanding further to ±4 gives no consistent additional benefit. The default ±1 has the best IQ, DD, and LPIPS and ties for the best MS, which supports it as a conservative choice.
Both extensions act only during offline calibration: they keep valid quantization codes and fixed row radii, and change neither the deployed representation nor the inference operations. The extra cost is small: correction takes 12.05 s for the default k=2,±1, 12.21 s for k=4,±1, and 13.74 s for k=2,±4. However, the gains are at most 0.03 dB PSNR and mixed across metrics, so we keep the default.