- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes PixelSTR, a lightweight distillation framework for text image super-resolution that aims to transfer image restoration priors from a large diffusion model to a non-diffusion feed-forward network, thereby reducing inference costs. Specifically, the method employs a frozen diffusion model as the teacher and trains a lightweight NAFNet student through RGB and spatial-gradient matching, together with reconstruction, perceptual, and OCR supervision based on ground-truth high-resolution images and text annotations. Experimental results demonstrate that PixelSTR achieves approximately a 15.8× forward-pass speedup while using only 0.98% of the teacher's restoration-module parameters, maintaining competitive image restoration quality and text recognition performance.
Critical strengths
-
Comprehensive experiments demonstrate a favorable efficiency–performance trade-off. PixelSTR achieves a 15.8× forward-pass speedup using only 0.98% of the teacher's restoration-module parameters, while maintaining competitive image restoration quality and text recognition performance.
-
Simple and practical framework design.
Critical weaknesses
-
It remains unclear why completely abandoning the diffusion is a necessary or sufficiently well-motivated design choice, rather than further compressing and accelerating diffusion models themselves. The authors should provide a clearer justification for this fundamental design decision and explain the distinct advantages of their approach over retaining a compact diffusion-based architecture.
-
Why do RGB teacher supervision and gradient teacher supervision, which appear to be two relatively independent loss terms, yield almost no improvement when applied individually, yet lead to substantial performance gains when combined? Although the paper demonstrates this phenomenon experimentally, it does not provide a sufficient explanation for the underlying mechanism.
-
Limited novelty. The main contribution of this work lies in distilling image restoration priors from a large diffusion model into a lightweight, non-diffusion feed-forward network, thereby significantly reducing inference costs while maintaining competitive restoration performance. However, prior work DTKD [1] has already investigated cross-architecture Diffusion-to-Non-Diffusion distillation for image super-resolution, transferring generative priors from a diffusion model to an efficient Transformer student through frequency-domain knowledge distillation. Therefore, the authors should more clearly articulate the key technical innovations of their method compared with DTKD.
[1] DTKD: Diffusion-to-Transformer Heterogeneous Knowledge Distillation for Efficient and Perceptually Enhanced Super-Resolution
Questions for authors
Please see weakness section.
Response to Reviewer xehb
W1. Why deploy a feed-forward student? At 8.62M total parameters, PixelSTR gains 3.8720 dB and 13.69/15.77 pp in TransOCR/PaddleOCR (T/P) over pruned TOLA on CTR (n=8,089). L2 channel pruning narrows the denoiser, VAE, OCR and MoM. With the intact latent refiner, the count is 8.624603M, 16 fewer parameters than PixelSTR. Pruned TOLA retains single-step diffusion and uses TOLA’s recipe for 50K recovery updates. PixelSTR uses GT and RGB/gradient KD for 50K steps and leads TinySR on all five metrics.
Efficiency. PixelSTR uses the fewest parameters and MACs and runs 15.8× faster than TOLA.
Real images. PixelSTR leads bicubic, OSEDiff and TinySR on RealCE-1K (1,000 crops, MARCONet++ [1]) and gains 1.0019 dB over OSEDiff on CTR.
W2. Why do RGB and gradient targets help together? Together they constrain stroke intensity and adjacent-pixel changes. Matching finite differences alone allows a constant offset. Matching both lowers edge, non-edge and image-gradient MAE by 10.48%, 31.71% and 16.60% versus GT alone. All models use the same HR edge mask per CTR image (top 25% by finite-difference magnitude).
Joint KD improves word recognition. It reads 630 strings both single-term students miss and corrects 749 GT-only errors (net gain 542). The joint gain exceeds the sum of the individual gains by 1.4299 dB and 6.57 pp in TransOCR accuracy.
The joint student has better-aligned KD gradients. Their mean cosine is 0.4484 for the joint student and 0.0137 for RGB only. Median norm ratios differ from 1 by 0.03%–1.53%. On eight fixed batches of 16, we sum the RGB and 0.1-weighted finite-difference parameter gradients and divide the sum’s L2 norm by the RGB-gradient norm.
GT and teacher update directions conflict. A negative gradient cosine means teacher descent increases GT loss to first order. Of 128 inputs, GT opposes RGB on 18 (14.06%), finite differences on 29 (22.66%) and joint KD on 18 (14.06%) (Appendix I.2, Fig. 10, p. 22). RGB-L1 derivatives oppose where (S−H)(S−T)<0 (S student, H reference, T teacher), averaging 34.69% of channel positions. In Appendix I.2, Fig. 11, p. 22, this fraction is 36.23% and GT/joint-KD cosine is −0.1649.
W3. What does PixelSTR add beyond DTKD? PixelSTR distills stroke intensity and edges by matching teacher RGB and adjacent-pixel differences at corresponding locations (fixed weights 1.0/0.1). Transcripts supervise character identity. DTKD [2] targets perceptual quality in natural images, emphasizes mid-frequency wavelet bands and increases the KD weight during training.
CTR recognition and restoration. Both recognizers rank PixelSTR above the two wavelet controls in the CTR table. Against the fixed-weight control, the T/P gains are 1.38 pp and 3.74 pp, respectively. PSNR rises 0.8973 dB and LPIPS falls 0.1060. Fixed beats linear by 0.0981 dB and 0.57 pp P-ACC.
Controlled DTKD-style wavelet baselines. We keep TOLA, the 8.62M NAFNet student, CTR data, GT losses (L1, LPIPS, OCR), optimizer, batch 16 and 50K updates. Two-level Haar KD replaces RGB/gradient KD. Low/mid/high weights are 0.25/0.50/0.25. The KD weight rises linearly from 0 to 1 or stays at 1.
References