- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies aesthetic supervision for pixel-space text-to-image pretraining. The authors first compare ten perceptual encoders, find that language-supervised encoders perform well under a shared recipe, and replace DINOv2 with SigLIP. They then introduce GESSO, which distills absolute teacher score gaps between reference images and their one-step reconstructions into a per-patch cosine distance produced by a lightweight projector on frozen SigLIP features. After training, the projector is frozen and its distance becomes an additional generator training loss. The authors report improvements in automated preference metrics and FID at three model scales, supported by three matched B/32 training repeats, VLM pairwise judgments, and complete-pipeline results after supervised finetuning.
Critical strengths
S1. The method provides a practical improvement that is straightforward to integrate. In a pipeline already using a SigLIP perceptual loss, GESSO reuses feature computation, adds approximately 0.8% to training step time, and incurs no inference overhead. This gives it practical value as a training recipe (Section 3.3, Table 5).
S2. The gains over a stronger baseline are consistent and have some evidence of repeatability. In Table 3, GESSO improves B/32 HPSv2.1 from 21.12 to 22.03 and reduces FID from 10.25 to 9.63 relative to PixelGen-SigLIP. Trends across scales and automated scorers, together with matched branches from three independent B/32 training runs, make me confident that the complete recipe's gains are not solely due to the encoder replacement or generation seeds (Tables 3–4 and 10).
Critical weaknesses
W1. Attribution of the central methodological contribution remains insufficient. In Table 2, the scalar alternatives backpropagate through the original MUSIQ network, whereas the per-patch alternative backpropagates through a trained SigLIP projector, changing the backbone, distillation procedure, and loss formulation simultaneously. Better performance of the complete recipe therefore does not establish that the advantage comes from the reference-anchored per-patch distance rather than transferring the teacher signal into a more optimizable representation. A random projector does not replace a scalar student trained with the same teacher supervision. In addition, the degradation-only control removes other projector training terms, leaving aesthetic supervision insufficiently separated from general reconstruction-quality supervision(Sections 4.3 and 4.9, Tables 2 and 11, Appendix B.3).
W2. The evaluation design limits confidence in generalization and the interpretation of preference gains. Appendix A.1.5 states that the teacher, switch-on step, and weight were selected on the reported benchmarks, without clearly establishing whether configuration selection and final evaluation used independent samples. Separately, VLM win rates include only pairs judged consistently after swapping image order, without complete win, loss, and exclusion counts. The judging prompt also prioritizes caption adherence and realism, so these results represent overall preference on that subset. Agreement across automated metrics is valuable, but extrapolation to human aesthetic preference should remain limited without direct human validation(Section 4.7, Appendices A.1.5 and B.6).
Questions for authors
Q1. Regarding W1: Are results available for scalar-score or global-distance students sharing the SigLIP backbone and teacher supervision, or degradation-supervision controls that preserve the auxiliary losses? If not, which existing evidence separately supports the independent roles of per-patch structure and aesthetic supervision? Please delimit the conclusions supported by the current experiments.
Q2. Regarding W2: Which prompts, samples, and metrics were used to select the teacher, switch-on step, and weight? Which final results were entirely excluded from configuration selection?
Q3. Regarding W2: Please provide each VLM's win, loss, and order-inconsistency counts for GESSO versus PixelGen-SigLIP. Do the conclusions hold when inconsistent judgments are treated as ties? Is human preference agreement already available for outputs from the models in this paper?