acceptodds
Under review as a conference paper at ICLR 2027

Guided Image Enhancement Path Diffusion Model

Abstract

Text-to-image diffusion models generate coarse structure and fine details within a single denoising process, leaving their coarse-to-fine hierarchy only implicitly modeled. We introduce GIEP (Guided Image Enhancement Path), which decomposes an image into ordered coarse-to-fine components and models their conditional dependencies. We show that the score of each partial reconstruction can be expressed as a sum of reusable per-level score contributions. This result yields a cumulative score-matching objective that supervises all levels jointly in a single shared network. Causal attention along the path axis enforces dependencies from coarse structure to fine detail, while level-specific text conditioning guides different levels of visual granularity. Experiments demonstrate leading perceptual quality and competitive compositional fidelity at a substantially smaller parameter scale; human evaluators also prefer GIEP for detail realism and saliency.

Primary area:
generative models
Keywords:
Text-to-Image Generation, Diffusion Models, Spectral Decomposition

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tYuwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GIEP, a text-to-image diffusion framework that adds an explicit coarse-to-fine path dimension. Images are decomposed into ordered frequency components, and their cumulative reconstructions are jointly denoised by a shared network. The method combines causal path attention, level-specific text conditioning, and a cumulative noise-prediction objective. The 4B model is trained from scratch on a 500M-image corpus and achieves competitive perceptual quality and a GenEval score of 0.86.

Critical strengths
  1. Making levels of detail explicitly addressable provides a clear architectural approach to modeling coarse-to-fine dependencies. Weight sharing allows the levels to interact without requiring separate models for each stage.
  2. The evaluation covers perceptual quality, compositional alignment, and human preferences. The component ablations are useful, particularly the causal-mask ablation, which supports the benefit of directional interactions between levels.
  3. Appendix G clearly reports the computational tradeoff. Relative to K=1, K=5 requires approximately five times the training FLOPs and 3.7 times the inference latency. The limitations section appropriately distinguishes parameter efficiency from computational efficiency.
Critical weaknesses

W1. The benefit of the enhancement path is not established under equal training compute.

Table 3 improves CLIPScore from 31.66 to 32.81 and FID from 25.03 to 22.61, but Appendix G reports 4.99 times the training FLOPs per sample for K=5. Figure 6 compares progress by epoch rather than cumulative FLOPs. These results therefore do not establish whether the path design outperforms a longer-trained K=1 model under the same computational budget.

W2. The score-matching derivation does not account for the predictor’s causal-prefix inputs.

The cumulative predictor receives the entire causal prefix, whereas Eq. 7 identifies its target with the marginal score of an individual level. With independent per-level noise, denoising regression generally targets a prefix-conditioned score, and Appendix D does not justify equating it with the marginal score. This leaves the claimed theoretical justification for the objective and its use during sampling unresolved.

Questions for authors
  1. Can you compare K=5 against a longer-trained K=1 model with the same parameter count and total training FLOPs, and plot quality against cumulative FLOPs?
  2. How does the equivalence in Eq. 7 account for conditioning on the causal prefix, and which score is used to update each path state during sampling?
  3. Are prompts expanded at evaluation time for GenEval, HPSv2, COCO-30K, and the human study, and do baselines receive the same expanded prompts?
Reviewer_4E5BOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

GIEP decomposes images into coarse-to-fine components and jointly denoises cumulative partial reconstructions using a shared network. The model introduces causal attention along the path dimension and level-specific text conditioning. Its theoretical argument claims that reusable atomic score contributions yield a cumulative score-matching objective equivalent to marginal Fisher divergence at each level. The system is trained from scratch on a 500M-image corpus, and the paper reports strong perceptual and compositional results, human preferences, and smaller-scale structural and compute ablations.

Critical strengths
  1. Explicit coarse-to-fine generation is an interesting modeling direction. The shared network, causal path attention, and level-specific text conditioning provide a concrete architecture for studying structured generation. This could have value even if the present theoretical interpretation needs revision.
  2. The reported empirical results and ablations are promising. GIEP obtains 0.86 GenEval versus 0.87 for FLUX.2-dev, with reported human preferences for fine-detail quality. Ablations examine the causal mask, text conditioning, decomposition, and path depth; Appendix G also discloses the additional compute cost. These provide evidence that the empirical design merits investigation.
Critical weaknesses
  1. The multi-level design incurs substantial inference overhead, and its quality–efficiency trade-off is not sufficiently convincing. Maintaining five path levels throughout the 75-step denoising process increases computation and activation memory despite weight sharing. Appendix G, Table 6 already shows latency rising from 20 to 74 seconds per image and peak inference memory from 14 to 33 GB when moving from K=1 to K=5, measured at 1024×1024 on one A800 with BF16 and batch size 1. In the accompanying 1M-image ablation, CLIPScore improves from 31.66 to 32.81 and FID from 25.03 to 22.61. The 3.70× latency and 2.36× memory costs are substantial, and the paper does not establish whether comparable quality could be achieved more efficiently with fewer levels or an alternative allocation of computation. A smaller parameter count alone does not demonstrate practical efficiency.
  2. The connection between the formulation and the generation procedure remains difficult to follow. The roles of frequency components, cumulative images, and network predictions are not brought together in a sufficiently clear account of inference. In particular, the relationship between the theoretical scores and the network's access to coarser noisy levels needs clarification.
Questions for authors
  1. Inference efficiency: Building on Appendix G, could you report end-to-end inference latency, throughput, and peak GPU memory for the full evaluated system alongside representative baselines under a consistent hardware and image-resolution protocol? Please specify whether prompt expansion is included and show the generation-quality versus inference-cost trade-off for different path depths, so that the practical benefit of using five levels can be assessed.
  2. Method clarification: Could you provide a simple example showing how the path-level predictions update the image states during inference and how this procedure corresponds to the mathematical formulation?
1 comment

The authors have not provided a response. Therefore, I will maintain my original negative assessment of this work.

Reviewer_N1N6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces GIEP, a text-to-image diffusion framework that decomposes images into ordered frequency components and jointly denoises their cumulative reconstructions.

Critical strengths

None.

Critical weaknesses
  1. The paper is hard to read. The use of language is strange. The citation is very misleading. A lot of the claims are not supported by the cited paper at all, discussed below. The citation format is also done incorrectly (should to use \citep to include parentheses). The related work section is filled with not-so-relevant work. The equations and notations in the method section is overly complex to describe supposedly simple things, making it impossible to understand.

  2. The motivation of the paper is not substantiated. It claims that traditional diffusion models with a single denoising schedule has implicit coarse to fine generation process, which causes negative entanglement and degrades quality. This claim is not substantiated by the paper nor is supported by the cited paper.

  3. Section 3 theory is also incorrect. The model predicts the score for level k using not only the noisy image at level k, but also all earlier noisy levels. Those earlier levels provide extra information. Therefore, the optimal prediction is a score conditioned on the earlier levels, not the marginal score of level k claimed in Theorem 3.1.

  4. The method is underspecified in the paper. The paper says that inference uses 75-step DDPM (quite a strange choice nowadays), but it does not provide an algorithm specifying (1) whether levels are updated jointly or sequentially, (2) which score is used for each update. (3) how the independently corrupted training prefixes correspond to inference-time states, (4) why this procedure samples the claimed joint path distribution.

  5. The evaluation is completely not standard. The model is trained on a private 500M-image high-resolution corpus with an image-aware VLM caption-expansion pipeline. The paper does not separate gains from the proposed architecture, as opposed to training data quantity and quality. It is also unclear whether competing systems receive the original prompts or the same expanded prompts. Comparing complete systems under their “recommended configurations” cannot establish that GIEP’s gains arise from the proposed method.

  6. If I understand correctly, having additional path also incur additional training and inference computation. The computational efficiency aspect is not discussed and compared at all.

Questions for authors

Please address the weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.