acceptodds
Withdrawn from ICLR 2027

Flat Features, Detailed Images? Simple Detail Control for Diffusion Transformers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Prompt engineering is commonly used to encourage text-to-image models to generate more detailed images. However, it does not provide reliable, continuous, or disentangled control. We analyze feature tokens produced with and without detail-enhancing prompts across six diffusion transformers and observe a common pattern: these prompts tend to increase the entropy of each token's energy distribution while largely preserving its ordering. That is, flattening the distribution increases the level of detail. Motivated by this counter-intuitive finding, we introduce a simple training-free approach that continuously adjusts the level of detail according to a scalar. Specifically, it shifts the squared channel magnitudes of each token toward a higher entropy distribution with a time-aware scheduler to concentrate the intervention at intermediate sampling steps. Across extensive experiments on different types of diffusion transformers, we show that our approach provides continuous control over detail metrics while largely preserving image quality and semantic structure. It requires only a few lines of code and adds negligible inference time overhead.

Primary area:
generative models
Keywords:
Text-to-Image Generation, Training-free Control, Diffusion Transformers

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_JZm7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how detail-oriented prompt suffixes affect feature distributions in diffusion transformers. It observes increased entropy in squared channel magnitudes, often with largely preserved channel rankings, and proposes Token Flattening to reproduce this behavior directly. The intervention redistributes channel energy while preserving token norms and signs, with a noise-dependent schedule controlling its timing. Experiments evaluate detail statistics, image quality, structural changes, and comparisons with several training-free guidance methods across multiple backbones.

Critical strengths
  1. The method is simple and computationally inexpensive. It provides a continuous control parameter without additional training or network evaluations. On FLUX.2 Klein 9B Base, the reported inference overhead is approximately 3%, with an HPSv3 improvement and nearly unchanged CLIP alignment.

  2. The feature analysis provides an informative empirical observation. Comparing identical latent states under different text conditions helps isolate the effect of the selected suffixes. The content-extension control and additional-backbone experiments strengthen the analysis, while the appendix transparently documents weaker results and failure cases.

Critical weaknesses
  1. The central evaluation does not adequately establish meaningful detail improvement. Edge strength, Laplacian variance, spectral edge frequency, and wavelet energy can also increase through sharpening, noise, or artifacts. HPSv3 offers partial support, but Qwen’s best positive-setting improvement is only 12.011 to 12.013 without uncertainty estimates. Downsampled LPIPS and global DINO features do not resolve this local-quality concern.

  2. Reliable, structure-preserving control is not sufficiently characterized. Table 3 compares one operating point per baseline, rather than the relevant detail–quality–structure trade-off. Likewise, monotonic dataset averages do not show whether individual prompts respond consistently. These limitations matter because predictable control, rather than merely changing image statistics, is the method’s main practical claim.

Questions for authors
  1. Does the method outperform sharpening or local-contrast enhancement at matched structural deviation in a blinded, full-resolution assessment of natural detail and artifacts?

  2. Can you report short strength sweeps and the fraction of individual prompts showing the expected detail change without material layout changes?

Reviewer_SdVtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper observes that detail-enhancing prompts tend to flatten token energy distributions in diffusion transformers. It uses this observation to propose a training-free method for adjusting image detail through a scalar, evaluated across several models and against existing guidance methods.

Critical strengths
  1. The observation is interesting and counter-intuitive, supported by substantial analysis across models.
  2. The results are appealing and convey the method well, including the comparisons to baselines. I appreciate that the method is training-free, and the added inference time is small.
Critical weaknesses
  1. The motivation for this control is not sufficiently clear to me. The authors haven't discuss the application where such control can contribute. Also, if I want a more detailed image, why not ask for it in the prompt? A direct comparison with tuned or LLM-assisted prompting is missing. Furthermore, the introduction criticizes prompting as unpredictable and not localized, but the proposed control is also model-dependent and global, with structural and appearance changes acknowledged in the appendix. Besides, increasing detail metrics does not establish more useful or precise control, making this motivation somewhat misleading.
  2. The observation is interesting, but I am not yet convinced that the demonstrated contribution is sufficient. The explanation of the observed phenomena remains speculative and the application is limited to global detail adjustment. Other generation properties, image editing, or a concrete application where this control is useful would make the broader significance more convincing.
  3. FlashAttention is not supporting saving the attention maps, which raise the concern that the proposed method might be slower then vanilla denoising with FlashAttention enabled. I would appreciate such comparison and clarification.
Questions for authors
  1. Both changing the prompt and changing the scalar require regenerating the image. In which settings does the scalar offer a meaningful advantage to the user?
  2. Beyond the proposed intuition, is there evidence explaining why the effect occurs across models? Have the authors checked other vision transformer models beyond generative models? Have you explored properties beyond detail or regional control?
Reviewer_fqb2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper analyzes how detail-enhancing prompt suffixes change intermediate tokens in diffusion transformers. Across six DiTs, these suffixes slightly increase the entropy of each token’s normalized squared-channel distribution while mostly preserving channel rank. Based on this, the authors propose T$\flat$, a training-free plugin that interpolates each token’s energy distribution toward uniform inside the residual branch, with strength set by a scalar α and a time-aware scheduler that peaks at intermediate noise levels. On FLUX.2 Klein 9B (base and distilled) and Qwen-Image, varying α\alpha changes four low-level detail metrics monotonically with small CLIP changes, and T$\flat$ beats several guidance-based baselines at about 3% latency overhead.

Critical strengths
  • Simple and cheap. The method is a few lines of code, needs no training, and adds about 3% latency, versus about 1.5× for guidance-based alternatives (Table 3).
  • Monotone control in both directions on the three main models (Table 2), including detail reduction at negative α\alpha.
Critical weaknesses
  • The motivating alternative, prompt engineering, is never compared quantitatively. The paper’s case for T$\flat$ is that detail-enhancing prompts are unreliable, non-continuous, and alter structure. Yet Table 3 has no prompt-based baseline, e.g., the paper’s own searched suffixes, graded variants (“detailed”, “highly detailed”, “ultra detailed”), or an LLM prompt rewriter.
  • The detail metrics don’t measure “detail” in the sense users care about. All four metrics are low-level frequency statistics, which also increase with noise, aliasing, and artifacts, and the paper’s own failure cases show “sharp artifacts” and “unnatural background clutter” at larger α\alpha.
  • No human study measures whether users perceive more meaningful detail at equal quality, or whether α\alpha gives perceptually even steps. A small user study (perceived detail, quality, and preference across α\alpha) would be far more convincing than the frequency statistics.
Questions for authors
  • How does T$\flat$ compare quantitatively with prompt suffixes of graded strength, and with an unsharp-mask or high-frequency-boost post-process, on the same metrics plus HPSv3 and structure distance?
  • How does T$\flat$ compare with Concept Sliders [1], which also provides continuous, structure-preserving attribute control via a slider scale?
  • Can you provide a user study on perceived detail and quality across α\alpha, compared with baselines?

Reference:

[1] Rohit Gandikota, Joanna Materzyńska, Tingrui Zhou, Antonio Torralba, David Bau. "Concept Sliders: LoRA Adaptors for Precise Control in Diffusion Models" in European Conference on Computer Vision (ECCV 2024).

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

Related papers

Discussion (0)

Sign in to comment.