- Rating
- 2
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper targets the per-step output discrepancy of a post-training-quantized diffusion denoiser under DDIM-family sampling. With the quantized backbone and DDIM schedule fixed, the authors propose SPEC, a 27.9K-parameter sampling-time corrector that adjusts the quantized denoiser’s time input at sparse steps and then predicts the remaining output error, trained in two stages on recollected trajectories. For stochastic DDIM, t-VR additionally fits a Student-t to the post-correction residual and subtracts its estimated variance from the injected Gaussian noise.
Critical strengths
S1. The decomposition of quantization error into a predictable same-state component and a post-correction residual, measured by PER, gives a clear diagnostic of what the corrector actually removes.
S2. Training the error branch on trajectories recollected after the temporal branch is active is a sound design choice that aligns training states with deployment states.
S3. The corrector is lightweight (27.9K parameters, 0.36 GPU-h offline on CIFAR-10) and leaves the quantized backbone and DDIM schedule untouched, so it is orthogonal to backbone-side PTQ methods.
Critical weaknesses
W1. (Marginal and statistically uncertain gains.) The improvements over Q-Diffusion on the same backbone are small relative to the reported noise: 4.89→4.81 (±0.03) on CIFAR-10 W4A8 at η=0, 3.97→3.90±0.02 and 5.55→5.49±0.05 at η=0.5, and 10.78→10.63 on LSUN-Bedroom W8A8 with only 3k samples, while LSUN-Church W8A8 regresses (7.50→7.56). Moreover, on CIFAR-10 the quantized model already has lower FID than FP32 (3.80 vs 4.22 at η=0; 4.31 vs 5.78 at η=1), so FID cannot tell whether SPEC recovers the full-precision behavior it is trained to match. Since the method’s stated objective is fidelity to , Table 1 should report distance-to-FP32 metrics (e.g., LPIPS or MSE between SPEC and FP32 samples generated from the same initial noise), along with multi-seed results and larger sample counts on LSUN.
W2. The DPM-Solver++ transfer test is too thin to support any claim about the method’s scope. Appendix C.2 applies the DDIM-trained SPEC to deterministic third-order singlestep DPM-Solver++ without retraining, and FID roughly doubles at both W8A8 and W4A8. The authors hypothesize stage-wise residual coupling and a state-distribution mismatch as the cause, but provide no diagnostic to test it and do not explain how the DDIM-grid refresh set and integer time offsets are mapped to solver stages. Although they state that solver-matched training would be needed, they never run it, even though the staged training recipe is solver-agnostic and costs under 1 GPU-h. It therefore remains unclear whether the failure is a design limitation or a training-distribution mismatch. The few-step multistep regime that matters most for low-bit PTQ is also untested, and since t-VR requires stochastic sampling, the full pipeline is demonstrated only for stochastic DDIM.
Questions for authors
Q1. For readers’ ease of following, please define in equation form (i) the SmoothL1 function in Eq. (8), including and the element reduction, and (ii) the “Relative Error Energy” on the y-axis of Figure 1, including which states are pooled and how it relates to PER in Eq. (4). On (i), with the default and discrepancies mostly far below 1, SmoothL1 reduces to MSE/2, so please clarify the chosen and why SmoothL1 was preferred over MSE.
Q2. What is the performance when the state-conditioned temporal branch is replaced by a per-timestep constant offset lookup table? Please also report the distribution of the learned oracle offsets across timesteps, including their sign, magnitude, and variance. If the within-timestep variance is small, state-conditioning would appear unnecessary; we would welcome the authors’ view on this.
Q3. In Table C-1, training the error branch with the MSE objective alone reduces IS from 9.36 to 8.54, below the uncorrected Q-Diffusion baseline. What does the corrector sacrifice in this setting, and why does adding the directional and magnitude terms recover IS while leaving FID nearly unchanged?