- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes SelectDiff, a training-free method for subject-consistent text-to-image generation. Instead of sharing all cross-image attention tokens, each query attends only to its top-K most similar tokens in the other images. NSVA adapts K across denoising steps with an Otsu-style variance criterion, and FBS restricts cross-image attention to likely foreground queries. On ConsiStory+ with SDXL, the method reports better consistency, prompt alignment and speed than prior training-free approaches, and it extends naturally to multi-subject scenes.
Critical strengths
- It is a simple and elegant idea. Using query–key similarity itself as an implicit subject mask is a neat insight. It removes ConsiStory's explicit mask extraction while being plug-and-play and faster. The adaptive-K view based on between-class variance is also a creative angle.
- The method gives the best training-free CLIP-T, CLIP-I and DreamSim scores on SDXL. It handles multiple subjects without any special design (Fig. 4), and the authors honestly report a failure case.
Critical weaknesses
- Recent training-free methods, especially DiT-based ones (e.g. CharaConsist), are missing. The FLUX.2 row is compared only against SDXL baselines and is described as "reserved".
- NSVA is never compared quantitatively with the best fixed K or with the exact Otsu solution, and Fig. 6 is a single qualitative prompt. Table 2 includes an "FBS without SGSA" row, although FBS is defined through top-K scores.
- Key hyperparameters (η, Q-rate, affected layers and timesteps, sampler settings) are missing.
Questions for authors
- How is FBS implemented without SGSA (Table 2)?
- Could you add a quantitative comparison of NSVA vs. the best fixed K vs. exact Otsu, and a ConsiStory-style baseline on the same FLUX.2 backbone?
3 comments
Thank you to the authors for the detailed response and for providing the additional experimental results and implementation details. Overall, I find the rebuttal satisfactory and will update my score to 3.
Thank you for your careful review and for updating your score. We truly appreciate your constructive feedback, which has helped us substantially improve the paper.
Response to Reviewer JE2v
W1 and Q2. Recent DiT methods and the “reserved” wording
Thank you for this suggestion. We have added CharaConsist, a recent DiT-based method, and broadened the evaluation with aesthetic and VLM-based scores. Aesthetic quality is evaluated using LAION Aesthetic Predictor V2, with raw prediction scores reported. GPT-6 provides a custom composite assessment of text–image semantic alignment, visual quality, cross-image subject consistency, and background leakage. Higher scores indicate better overall performance across these dimensions, rather than an independent improvement in every component. We have included the evaluation prompts, model version, scoring scale, and aggregation procedure in the revised manuscript.
Table R1. Comparison with a recent DiT-based method.
SelectDiff+ improves CLIP-T and CLIP-I by 1.6 and 2.9, reduces DreamSim by 9.3, and obtains higher aesthetic and VLM scores. Runtime decreases by 5.1 seconds and memory by 14.7 GB. These results extend the system-level comparison to a recent DiT-based consistency method, and we have incorporated them into the revised manuscript. Sharing the DiT architecture category does not by itself establish identical backbone checkpoints.
We also apologize for the retained “reserved” wording. The experiments were completed, but the placeholder was inadvertently left unchanged during manuscript preparation. We have replaced it with: “The SelectDiff (FLUX.2) row in Table 1 reports results obtained under the ConsiStory+ protocol.”
W2 and Q2. NSVA versus the best fixed K and exact Otsu
Thank you for recommending these comparisons. Exact maximization of the Otsu-style between-class variance is possible for a given set of attention scores. However, this is a single-step optimum, not a global optimum for the denoising trajectory or final generation quality, because the score distribution changes over time.
NSVA reduces the search overhead of adapting K: it evaluates the objectives at K and K+1, forms a local finite-difference update, and applies the update at the next denoising step. Exact maximization instead evaluates all candidate K values. After sorting, prefix sums allow this evaluation in O(L) additional work per score vector, whereas NSVA requires two objective evaluations once the necessary statistics are available. These counts exclude similarity computation and acquisition of the statistics; the following experiment reports generation quality and measured runtime.
Table R2. K-selection comparison using the same cache settings.
Compared with the validation-selected fixed K, NSVA increases CLIP-T and CLIP-I by 4.9 and 5.1, reduces DreamSim by 3.9, and increases runtime by 2.3 seconds. Compared with exact Otsu, its three quality metrics are numerically better, while runtime is reduced by approximately 46.7%. These results support its quality–runtime trade-off; better final quality does not imply more accurate optimization of the single-step Otsu objective. Fig. 6 illustrates parameter effects and is not used as evidence that NSVA outperforms every fixed-K setting.
W3. Missing hyperparameters and implementation details
Thank you for pointing out the missing implementation details. The configuration described here uses η=1, a Q-rate of 0.8, application to all self-attention layers, 20 denoising steps, and the DDIM sampler. We have provided the complete settings and their backbone-specific scope, including the affected denoising timesteps, to improve reproducibility.
Q1. How is FBS implemented without SGSA in Table 2?
Thank you for identifying this ambiguity. As described in §4.3 on page 8, removing SGSA replaces sparse aggregation with naive cross-image attention fusion. In the FBS-only configuration, we still compute foreground scores from the mean top-K similarities and select the top-Q queries. These selected queries then attend to all reference keys and values, without restricting cross-image connections to the top-K subset. Unselected queries use within-image self-attention only.
Thus, this configuration retains FBS foreground-query selection while removing sparse cross-image feature aggregation. We have clarified this implementation in §4.3 and explicitly distinguished the top-K operation used for foreground scoring from the top-K restriction used for feature aggregation.