- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes PT-Flow one-step generative modeling framework, which utilizes Schrödinger bridge to construct analytical target generation function. They parameterize the data-side potential to approximate data distribution and the proposal mean to reduce estimation variance and serve as generator. The experiments are primarily conducted on ImageNet 256 × 256 and get good results.
Critical strengths
- It's quite interesting to explore generative modeling paradigms beyond diffusion/flow-based models, so any exploration in this direction should be encouraged.
- This paper presents detailed derivation of PT-Flow and gives substantial complexity analysis, which is novel and interesting.
Critical weaknesses
- Although PT-Flow parameterize proposal to reduce estimation variance, training this one-step generative model seems potentially unstable: there is no theoretical guarantee for the alternating optimization scheme will converge to optimal solution, and this type of two-player optimization is to be challenging to train such as GAN.
- The experimental evaluation is limited to ImageNet and the results is not sufficiently convincing to demonstrate the general effectiveness of PT-Flow.
Questions for authors
- I have generally understand the design principle of proposal, but it seems just a heuristic method and no guarantee for reducing importance weight variance. Can you give some theoretical evidence for this?
- I do not understand the statement in lines 195–197. How do you calculate "0.14" and what is the meaning of "the unit data scale"? It's quite confusing to encounter this statement at this place.
- In section 4, why do you present the ablation study first then show the main results? I think this ablation study should not shown in the main text because this is just how you tune the hyperparameters.
- As you state that "training is expensive", you should report the training time of your main experiments and compare it with other baselines.
Response to Reviewer LFxk
Thank you for asking what the proposal guarantees and what evidence supports the alternating training procedure. We address these questions separately below and report a completed second-dataset experiment.
W1/Q1. First, the potential network uθ and the proposal network mη do not have opposite goals as in a GAN. The potential network learns the bridge potential. The proposal network predicts the center of the importance proposal during training and serves as the one-step generator at inference. During each generator update, the potential is held fixed, and the generator is trained by regression to satisfy the proximal condition (Eq. 15). The generator loss does not use the value of the potential loss (Appendix I.2). A better picture is a half-Sinkhorn step: one side of the bridge has a closed-form expression, and the other is fit by gradient steps. The potential changes between updates, so the two training steps are still coupled.
Second, the variance fix is not just a heuristic. Theorem 3.9 gives an O(ε) bound on the log-weight variance of the centered diagonal proposal under condition (A6) and the stated remainder bound. This is a small-ε result. At our ImageNet setting, ε=0.01 and εd≈41, outside that regime. Table 7 checks how the proposal works at this setting: on the same frozen checkpoint, mean ESS/K is 0.0768 for the naive proposal, 0.9158 after recentering, and 0.9895 for the defensive mixture. Most of the observed gain comes from recentering.
On training stability, we completed the same 160,000-step schedule for the B/2, M/2, and L/2 models. This shows that the alternating procedure could be trained at all three sizes. A general convergence guarantee for the trained networks remains open, as we discuss in Appendix I.4.
W2. Evaluation beyond ImageNet. On FFHQ-256, PT-Flow-B/2 achieved FID-50k of 2.12 at 1 NFE. We used the same 256×256 latent pipeline and generated 50,000 images with 1 NFE. We computed FID using the 60,000 FFHQ training images as the real-image reference. On ImageNet-1K, the 1-NFE FIDs were 1.73, 1.62, and 1.54 for B/2, M/2, and L/2, respectively.
Q2. Origin and interpretation of “0.14.” This line was too compressed. The 0.14 comes from Proposition 3.4, with the scale of the bound written as a fraction of the data scale. The proposition bounds the W2 distance between the distribution of the exact conditional-mean output and the bridge marginal by the square root of the expected conditional covariance trace. The Laplace approximation gives 1+λˉ2εd for this quantity.
“Unit data scale” refers to assumption (A5): we scale the data to unit variance per dimension, so the root-mean-square distance from the data mean is d. Dividing by d cancels the dimension:
d2εd/(1+λˉ)=1+λˉ2ε.
At ε=0.01, neglecting the curvature correction gives 0.02, about 0.14. This is an approximate scale for the distribution-level bound. We will rewrite the passage to show this calculation, define (A5) there, and make clear where we switch from an absolute distance to a relative one.
Q3. Order of the experiments. We agree. We moved the main results table (Table 3) before the ablation, and pushed the tuning details to the end of the section. (Another reviewer asked us to move the algorithm into the main text; we did both in the same revision.)
Q4. Training time. All three ImageNet models followed the same 160,000-optimizer-step schedule. The wall-clock estimates use the measured time per step on each configuration.
For L/2, the effective global batch was 1,024 images per optimizer step. Each GPU contributed 128 images across eight gradient-accumulation steps. Training from scratch took an estimated 24.1 days on eight H100 GPUs.