Diffusion-Classifier Synergy: Reward-Aligned Learning via Mutual Boosting Loop for FSCIL

We sincerely thank you for your careful review and encouraging recognition of our work. We will respond to the questions you have raised in detail below.

Q1: Inconsistent performance superiority across all datasets.

A1: Thank you for this insightful comment. We appreciate the opportunity to provide a more nuanced analysis of our model's performance.

We agree that a holistic evaluation across multiple metrics is crucial for FSCIL. While average accuracy provides a general overview, we argue that final session accuracy and forgetting rate are more critical metrics for evaluating true continual learning, as they measure a model's long-term viability.

On CIFAR-100, our method demonstrates superior knowledge retention. From session 1 to 8, ALFSCIL's accuracy drops by 22.71% (from 77.88% to 55.17%), whereas DCS's accuracy drops by only 20.75% (from 75.96% to 55.21%). Crucially, our performance in the final session is higher than ALFSCIL's, highlighting our model's enhanced ability to both learn new concepts and mitigate catastrophic forgetting.

Furthermore, on the miniImageNet benchmark, DCS significantly outperforms ALFSCIL across the board: our average accuracy is 3.34% higher (68.14% vs. 64.80%), and our final session accuracy is 4.68% higher (57.99% vs. 53.31%). On the fine-grained CUB-200 dataset, while SAVC's average accuracy (69.35%) is the closest to ours (69.73%), DCS achieves a higher final session accuracy (63.40% vs. SAVC's 62.50%) and shows a smaller accuracy drop from the first to the last incremental session, indicating better long-term learning.

The performance variation stems from dataset characteristics. CUB-200's fine-grained nature presents a unique challenge due to high inter-class similarity. We see it as a strength that DCS, with a unified framework and without dataset-specific hyperparameter tuning, achieves consistent and significant improvements on diverse datasets. This demonstrates the generalizability of our core concept. However, we agree that exploring reward designs tailored to the semantic distributions of fine-grained datasets is a valuable direction for future research.

We hope this clarifies the significant advantages of our approach. Thank you again for your valuable feedback.

Q2: Concerns on Computational Efficiency.

A2: The primary computational overhead of our method is concentrated in the image generation phase that precedes the training of each incremental session. It is important to note that our approach, by leveraging the training-free DAS algorithm, does not require any fine-tuning of the diffusion model itself.

In our experiments, the highly parallelizable image generation for a typical incremental session takes 5-10 minutes (30 images for each new class and 5-10 for each old class). The subsequent classifier training on this augmented data is very fast, taking less than a minute per session. Given the limited number of sessions in standard FSCIL benchmarks, we believe this time overhead is a reasonable and acceptable trade-off for the performance gains achieved.

Crucially, our method adds no overhead at inference time. The diffusion model is discarded after training, and the final classifier is a standard ResNet. Therefore, its architecture, parameter count, and inference speed are identical to the baseline.

Finally, the computational cost is flexible and can be further optimized by generating fewer samples, adjusting the DAS configuration, or employing a more lightweight diffusion model, allowing users to balance performance with available resources. We will add a detailed analysis of this in the final version of the paper.

Q3: Concerns regarding the fairness of comparison due to the use of high-resolution image generation.

A3: We would like to respectfully clarify a potential misunderstanding. While the images were initially generated at a $512 \times 512$ resolution, they were always resized down to the native resolution of the respective dataset (e.g., $32 \times 32$ for CIFAR-100) using torchvision.transforms.Resize before being input to the classifier.

Crucially, this downscaling process is inherently lossy; the resulting low-resolution image does not contain more information than an original image of the same size and is simply a synthetic sample that conforms to the standard data format. Therefore, the model was not trained on high-resolution images in any form, ensuring no unfair informational advantage was conferred.

The decision to generate at a higher resolution was a practical necessity. The performance of the Stable Diffusion model we employed is optimized for specific resolutions. This is because its core model was pre-trained on images of larger dimensions, such as $512 \times 512$, $1024 \times 1024$, or even higher. At the very low native resolutions of the FSCIL datasets, the model is unable to generate coherent, semantically meaningful images, often producing only noise. Our approach allows us to use this state-of-the-art generative tool effectively, while still maintaining a fair and standardized evaluation protocol for the classifier.

Q4: Lack of visual examples and analysis for the images.

A4: We agree that visualizing the generated images is crucial for a complete understanding of our method's mechanics. We will certainly incorporate a dedicated section with visual examples and a comparative analysis in the final version of the paper.

Due to the format limitations of this rebuttal, we are unable to display images directly. Instead, we offer the following textual description of their key characteristics:

• Semantic Correctness and Style Diversity: Guided by both text prompts and our multi-faceted reward system, the generated images are not only semantically correct but also rich in scenic diversity. Since our diffusion model is guided by the classifier’s state rather than the original dataset's images, the generated samples exhibit a broader range of styles than the source data. This process, constrained by our logits-level rewards, ensures the classifier can correctly identify these novel variations, thereby significantly enhancing its generalization.

• Subject Amplification: A key difference from real images lies in the portrayal of the subject. Our method provides a clear semantic target, prompting the diffusion model to amplify the subject within the frame, often resulting in a larger area and finer detail. While this helps the classifier focus on and learn richer class-specific features, it does differ from many real-world images where subjects can be small or part of a complex, multi-object scene. Although the content of the generated image can be constrained by designing richer text prompts, we believe that developing more adaptive algorithms to generate scenes more aligned with the original dataset is a promising direction for research.

• Nature of "Hard" Samples: For images generated to be intentionally confusing (e.g., an 'X' that fools the classifier into seeing a 'Y' via the R_CSCA reward), the results are particularly interesting. Visually, the generated image remains clearly identifiable as 'X' to a human observer and does not appear as a blend of 'X' and 'Y'. However, these images contain subtle, high-frequency patterns, akin to adversarial noise. As our reward signals confirm, these patterns are highly effective at misleading the classifier. We hypothesize this is because the diffusion model learns to exploit the classifier's high sensitivity to texture and fine details, which humans often overlook in favor of holistic shapes. A thorough investigation of this adversarial-like phenomenon is a compelling area for future work.

We sincerely thank you for your careful review and encouraging recognition of our work. We will respond to the questions you have raised in detail below.

Q1: Clarity on Technical Contribution Beyond DAS.

A1: Thank you for your positive feedback. This is an insightful question, and we appreciate the opportunity to clarify the core novelty of our work.

We agree that our method is built upon an alignment algorithm, for which we chose Diffusion Alignment as Sampling (DAS). As we state in the Preliminary section, having an effective, training-free method to align the diffusion model with a reward signal is a foundational premise for our framework. Indeed, our approach is not strictly limited to DAS. Any similar technique that achieves this goal, such as those proposed in [a] or [b], could be integrated into our system.

The central technical contribution of our work, Diffusion-Classifier Synergy (DCS), is the design of the mutual boosting loop itself. Our focus is on how to leverage the generated data distribution to create a controllable, joint co-evolutionary framework for a downstream task. The classifier's evolving state provides a rich, multi-faceted reward signal that guides the diffusion model to generate strategically beneficial samples, which in turn are used to refine the classifier's understanding. This concept of co-evolution is particularly crucial in extreme scenarios like FSCIL, where data is scarce and every new sample must be maximally informative.

In the future, we will certainly explore using novel distribution metrics and target alignment algorithms to optimize the sampling step. Thank you again for your valuable suggestion, which points toward an important direction for the future improvement of our work.

Ref:

[a] Cheng, Min et al. "Diffusion Blend: Inference-Time Multi-Preference Alignment for Diffusion Models."

[b] Li, Xiner et al. "Dynamic Search for Inference-Time Alignment in Diffusion Models."

Q2: Lack of analysis of training time.

A2: The primary computational overhead of our method is concentrated in the image generation phase that precedes the training of each incremental session. It is important to note that our approach, by leveraging the training-free DAS algorithm, does not require any fine-tuning of the diffusion model itself.

In our experiments, the highly parallelizable image generation for a typical incremental session takes 5-10 minutes (30 images for each new class and 5-10 for each old class). The subsequent classifier training on this augmented data is very fast, taking less than a minute per session. Given the limited number of sessions in standard FSCIL benchmarks, we believe this time overhead is a reasonable and acceptable trade-off for the performance gains achieved.

Crucially, our method adds no overhead at inference time. The diffusion model is discarded after training, and the final classifier is a standard ResNet. Therefore, its architecture, parameter count, and inference speed are identical to the baseline.

Finally, the computational cost is flexible and can be further optimized by generating fewer samples, adjusting the DAS configuration, or employing a more lightweight diffusion model, allowing users to balance performance with available resources. We will add a detailed analysis of this in the final version of the paper.

Q3: A detailed comparison between the proposed method and previous generative-based incremental learning methods.

A3: Thank you for your insightful suggestion. We agree that a detailed comparison with related generative-based methods would strengthen our paper, and we will incorporate this discussion into our final version. We hope this clarifies the distinctions and highlights the specific contributions of DCS in the challenging FSCIL context.

We would first like to clarify a key distinction in the problem settings. The cited works, DDGR and DiffClass, primarily address the general Class-Incremental Learning (CIL) problem. In CIL, while the data for new tasks may be limited, it is relatively sufficient, and the main goal is to mitigate catastrophic forgetting. Our work focuses on the more challenging Few-Shot Class-Incremental Learning (FSCIL) setting, where new classes are introduced with extremely few samples (e.g., 5-shot). This not only exacerbates forgetting but also makes it incredibly difficult for the model to learn robust new concepts.

Due to this fundamental difference in task settings and evaluation protocols, a direct comparison of experimental results is not feasible. However, a comparison of the design philosophies is indeed valuable and reveals the unique contributions of our DCS framework.

---

Comparison with DDGR

• Interaction Mechanism
  • DDGR employs a direct, gradient-based intervention. It calculates the classifier loss's gradient with respect to the noisy image at each sampling step and injects this gradient to steer the generation. This is a form of micro-level, direct manipulation.
  • DCS utilizes a more flexible Reward-Aligned Learning paradigm. We design a multi-faceted reward function that assesses the strategic value of a generated sample based on the classifier's state. Then, we use the DAS algorithm to optimize the sampling process to maximize this reward. This is a macro-level, goal-oriented guidance.
• Guidance Signal Complexity This is the core difference and our key innovation.
  • DDGR's guidance signal has a singular objective: improving the classification probability of the generated sample.
  • DCS designs a sophisticated, two-level reward system that goes beyond simple accuracy.
• Strategic Goal
  • DDGR primarily aims for high-quality replay of past data for rehearsal.
  • DCS has a broader strategic goal. It not only replays old data but also performs crucial data augmentation for the new few-shot classes and, as mentioned, generates targeted hard samples to enhance the classifier's discriminative power. It provides data for "review," "preview," and "remedial lessons."

---

Comparison with DiffClass

• Core Paradigm

  • DiffClass adopts a "preventative alignment" approach. At the start of each incremental step, it fine-tunes the diffusion model to align the generated data distribution with real and historical data. The information flow is one-way: from the generator to the classifier.
  • DCS establishes a "co-evolutionary synergy." It creates a continuous, bidirectional feedback loop where the classifier's real-time state dynamically guides the diffusion model's generation, leading to a synergistic improvement of both models.

• Generation Mechanism

  This represents a fundamental distinction in the technical implementation of the two methods.

  • DiffClass: Employs a model fine-tuning approach. At each incremental stage, it leverages LoRA and theMulti-Distribution Matching loss function to actively train and modify the weights of the diffusion model. This process involves gradient computation and model parameter updates.
  • DCS: Adopts a training-free guidance methodology. Throughout the entire generation process, the weights of the diffusion model remain fixed and unchanged.

• Guidance Signal Complexity

  • DiffClass relies on a Multi-Distribution Matching loss, which is a macro-level control that aligns the overall statistics of data distributions.
  • DCS's reward system offers precision guidance. It operates at a granular level, considering not just distribution alignment but the strategic utility of each sample.

Dear Reviewer eG2t,

This is a friendly reminder to check the authors' rebuttal and adjust your rating if necessary. Thanks for your contributions to the NeurIPS reviewing process.

Thanks,

Your AC

We sincerely thank you for your careful review and encouraging recognition of our work. We will respond to the questions you have raised in detail below.

Q1: Inaccurate phrasing of performance claims and suboptimal table formatting.

A1: We will carefully revise the manuscript to use more precise and nuanced language when describing our model's performance. We will adopt your suggestion to bold the best-performing result in each column of our comparison tables.

Q2: Necessity and Interplay of the Multi-Component Reward.

A2: Our design is a deliberate, hierarchical system, not an arbitrary mix, and we will clarify this in the revision. The core idea is that different rewards are deployed to solve different, specific problems.

1. Foundational Rewards: The feature-level rewards, $R_{PAMMD}$ and $R_{VM}$, work in tandem to establish a necessary semantic foundation. As shown in Table 3 in the main paper, adding $R_{PAMMD}$ alone improves the final session accuracy by +1.24% over the baseline. Incorporating $R_{VM}$ further boosts this relative improvement to +1.86%, demonstrating their combined utility in ensuring generated images possess core class coherence and rich intra-class diversity.

2. Targeted, Mutually Exclusive Rewards: Our framework employs a task-dependent strategy based on the learning objective, as detailed in Algorithm 1 (Appendix):

   • For New Classes: The goal is to build a robust initial understanding. We use $R_{PAMMD}$ + $R_{VM}$ + $R_{RC}$ to generate high-confidence examples that firmly establish the new concept for the classifier.
   • For Old Classes: The challenge is preventing confusion with new classes. Here, we use $R_{PAMMD}$ + $R_{VM}$ + $R_{CSCA}$ to generate strategically challenging "hard" samples that specifically probe and strengthen the decision boundaries.
   • The data in Table 3 in the main paper validates this specialized strategy. The penultimate row, which uses only the $R_{RC}$-based strategy for all classes ($R_{PAMMD}$ + $R_{VM}$ + $R_{RC}$), achieves a relative improvement of +3.50% in the final session. However, the final row, representing our full approach that also incorporates $R_{CSCA}$ for old classes, demonstrates a significantly higher final-session improvement of +5.64%.

3. For a more detailed analysis of the conflict associated with $R_{CSCA}$, we have provided an in-depth discussion in our response to Q3.

Q3: On the potential risks of the confusion-seeking R_CSCA and its balance with consistency rewards.

A3: Thank you for raising this important point regarding the balance and stability of our reward system. The framework manages this balance primarily through two key mechanisms.

First, the influence of R_CSCA is explicitly controlled by the scaling factor γ in Equation (14). This parameter governs the sensitivity of the dynamic weights to the cosine distance, allowing us to regulate the degree of confusion in the generated samples. This provides a clear mechanism to balance the objective of R_CSCA with the feature-level consistency rewards. We will clarify in our revision that this parameter can also be made learnable to enable adaptive balancing.

Second, the diffusion model's generation process is always conditioned on an explicit text prompt describing the target class. This textual conditioning provides a persistent semantic guide throughout the denoising steps, ensuring that the final output does not deviate significantly from the target class's core concept, even as R_CSCA encourages exploration near decision boundaries.

We believe this is a very important direction, and we are keen to explore the precise impact of these hard-sample generation strategies on the classifier's decision boundary as a significant part of our future work.

Q4: Inadequate Ablation Study Methodology and Component Necessity.

A4: We agree that a more comprehensive ablation study is crucial for demonstrating the necessity of each reward component. To address this, we have conducted a "leave-one-out" analysis on the CIFAR-100 dataset. Regarding the combination of rewards, as pointed out in Q2, R_RC and R_CSCA will not be used at the same time, so we choose the combination of R_PAMMD + R_RC.

The results are presented in the table below:

As demonstrated in Table 1, removing any single component results in a performance decrease, confirming that all four components contribute synergistically to the final performance of DCS.

Q5: Limited Generality Verification of the Reward Design.

A5: We appreciate you raising this point regarding the generalizability of our reward design. We would like to emphasize that the performance of our full DCS model across three diverse benchmarks is in itself the most direct and powerful evidence of our method's generality. We understand, however, that an explicit component-wise analysis on each dataset would further strengthen this claim. To make this point more explicit and fully address your query, we are happy to provide the sequential ablation studies for miniImageNet and CUB-200 below.

As these tables clearly show, the performance gains from each reward component follow a consistent and positive trend across all three datasets.

Q6: Applicability of the reward mechanism to alternative generative models (e.g., GANs, VAEs).

A6: Our primary consideration for using diffusion models over other generative models is their significant, principled advantages for adaptation to few-shot sets. The adversarial equilibrium between a GAN's generator and discriminator is delicate. Fine-tuning on a small dataset can easily disrupt this balance, leading to mode collapse. By contrast, diffusion models are architecturally more resilient and offer more flexible adaptation pathways.

Furthermore, the core mechanism of Diffusion Alignment as Sampling (DAS) is fundamentally and inextricably linked to the iterative reverse process that is the unique hallmark of diffusion models. DAS operates by intervening within a Markovian sequence of noisy latent states ($x_T \rightarrow x_{T-1} \rightarrow ... \rightarrow x_0$). It guides this trajectory at each step by resampling candidate states based on a reward function, a process enabled by the step-wise nature of denoising.

This locus of intervention is architecturally absent in GAN and VAE frameworks:

• GANs employ a generator that performs a single, deterministic mapping from a latent vector $z$ to a final sample $x$. There is no intermediate, sequential state between $z$ and $x$ in which to apply iterative guidance at test-time. Any "guidance" would have to involve either optimizing the input latent $z$, a different and more constrained problem, or altering the generator's weights, which constitutes re-training, not test-time alignment.
• VAEs similarly use a decoder to perform a one-shot mapping from a latent code to a sample. The generation is not a sequential refinement process that can be guided step-by-step.

Therefore, the DAS algorithm is conceptually incompatible with the architectural paradigm of GANs and VAEs.

Q7: On the comparison against non-reward-guided generation.

A7: The requested baseline is, in fact, the 'Baseline' row in our ablation study (Table 3 in the main paper). This baseline utilizes the same diffusion model for augmentation but omits our proposed reward guidance. As the data shows, the baseline's accuracy degrades to 49.57% in the final session. In contrast, our full method, which includes all reward guidance components achieves a final accuracy of 55.21%. We acknowledge this was not explicitly stated and will clarify its role in the revised manuscript. Thank you for helping us improve the paper's clarity.

Q8: Regarding the feasibility and sensitivity of the encoder.

A8: Yes, our DCS framework is designed to be encoder-agnostic. It is compatible with any architecture that provides feature embeddings and classification logits, as these are the direct inputs to our reward functions.

Regarding sensitivity, our framework demonstrates notable robustness. This stems from two key factors. First, in the standard FSCIL setting, the encoder is well-initialized during the base session where data is ample, ensuring high-quality initial reward signals. Second, the diffusion model's generation is also guided by explicit textual prompts. This textual conditioning acts as a strong semantic anchor, ensuring the generated images maintain class identity even if the classifier's feedback is imperfect. This dual-guidance ensures the co-evolutionary process remains on a positive trajectory, making our framework less sensitive to the encoder's initial quality than one might expect.

We sincerely thank you for your careful review and encouraging recognition of our work. We will respond to the questions you have raised in detail below.

Q1: Ablation on Generation Quality Metrics.

A1: Thank you for this insightful suggestion. We agree that analyzing the impact of our reward components on generation quality metrics like FID and CLIP score is valuable for a deeper understanding of our method.

We would like to clarify our initial reasoning for not including these metrics in the main paper. Our work operates in a strict few-shot setting, where both the real and generated samples per class are extremely limited (e.g., 5-30 images). Under such data scarcity, metrics like FID and CLIP score are known to be highly unstable and can produce misleading results. For instance, FID suffers from unreliable statistical estimations and high variance, while CLIP score on a few samples may not be representative of the true distribution's diversity.

Nevertheless, to address the your valid point, we have conducted a new set of ablation studies as requested. We performed these experiments on CIFAR-100, miniImageNet, and CUB-200, focusing on the first class of the first incremental session. For each setting, we generated 30 images and repeated the runs to report the average scores.

Crucially, we must emphasize two points:

1. The absolute scores computed on such a small number of samples will naturally be worse (higher FID, lower CLIP) than those reported in large-scale generation literature. The key insight lies in the relative improvement across different reward settings.
2. Due to resource constraints during the rebuttal period, this analysis is on a subset of classes. We plan to include a more comprehensive analysis in the final version.

The results are presented in the table below:

Analysis of Results:

As the table shows, R_PAMMD significantly improves both FID and CLIP scores over the baseline by promoting semantic consistency and diversity.

Adding R_VM further enhances these metrics by ensuring the feature spread aligns with real data.

The inclusion of R_RC leads to the best CLIP scores. This is expected, as R_RC directly optimizes for classification confidence, thus pushing generated samples to be more semantically pure.

Most importantly, using the confusion-aware reward R_CSCA degrades both FID and CLIP scores. This result, while seemingly negative, validates our hypothesis. The explicit goal of R_CSCA is to generate hard, confusing samples at the decision boundary, which are by nature less aligned with the true, clean data distribution. This demonstrates that our reward system can strategically generate not just high-fidelity images, but also targeted, challenging samples tailored to the classifier's evolving needs.

We also wish to clarify that these rewards are not all used simultaneously for the same purpose, as detailed in Algorithm 1 (Appendix):

• For new classes, we use R_PAMMD + R_VM + R_RC. The goal here is to generate high-fidelity, semantically pure samples to quickly and robustly establish the initial concept representation.
• For old classes, we use R_PAMMD + R_VM + R_CSCA. The goal is to mitigate forgetting and reduce inter-class confusion by intentionally generating 'hard samples' that lie near the decision boundaries of the new classes.
• The penultimate row of Table 3 in the main paper shows the result of using only the R_RC-based strategy for all classes, while the final row demonstrates the performance of our full, specialized approach.

We hope this detailed analysis addresses your concerns and further highlights the significance of our proposed reward strategy.

Q2: Comparison with Non-Generative Statistical Calibration Methods.

A2: We will first articulate the fundamental theoretical advantages of a generative approach like DCS over a post-hoc statistical calibration method like [a]. We will then present a new set of experiments directly comparing DCS to [a] on the quality of the learned distributions, as suggested.

While Goswami et al. [a] present an elegant and effective method for improving classification, its scope is inherently limited to a post-hoc statistical adjustment. Our generative approach, DCS, operates on a more fundamental level.

• Static vs. Dynamic Learning: The calibration method [a] is static. It calculates new class prototypes and covariances and then uses them in a frozen, non-trainable classifier (like NCM or a fixed linear probe). It improves inference but does not and cannot improve the feature extractor or the classifier's decision logic itself. In contrast, DCS is dynamic. By generating actual image instances (x_gen), we create new training data that can be used to fine-tune the classifier via backpropagation. This allows the model to actively learn a more robust decision boundary, rather than just passively using a better-estimated statistic.
• Extrapolation vs. Interpolation: The calibration method [a] relies on interpolation. It assumes a new class's distribution can be reasonably approximated by a weighted average of base class distributions. This works well when new classes are semantically close to base classes. However, if a new class is truly novel and semantically distant from all base classes, the calibration will be based on irrelevant information, and the resulting statistics will be poor. DCS, powered by a strong diffusion model, has the potential for true extrapolation. Guided by our multi-level rewards, it can synthesize novel feature representations that are not a mere blend of old ones, offering a path to generalize to genuinely new visual concepts.

In summary, calibration [a] is an excellent inference-time enhancement, while DCS is a comprehensive training-time framework that improves the model's core representation and decision-making capabilities.

To empirically substantiate this, we followed your suggestion to directly compare which learned distribution is closer to the real data.

• Experimental Setup: We calculated the FID to measure the similarity between the real test data and the data distributions produced by both methods. The experiment was conducted on the 10 new classes from the first incremental session of the CUB-200 dataset.

  • For our DCS method, we generated images for each new class and calculated the FID directly between the paths of these generated images and the real images using a standard FID implementation (e.g., PyTorch-FID with InceptionV3 features).
  • For method [a], since it operates purely in the feature space and does not generate images, a direct image-based FID is not possible. To perform a fair comparison, we sampled feature vectors from its calibrated Gaussian distribution and compared them to the real features.

• Results: The average FID scores across the 10 new classes are:

  • Method [a] (C-RanPAC, Feature-FID): 305.68
  • Our Method (DCS, Image-FID): 197.2

The significantly lower FID score for DCS provides quantitative evidence that our reward-aligned generative process produces a distribution that is a more faithful approximation of the true data distribution.

We believe this result, combined with our conceptual arguments, supports the value of our generative approach. However, we also wish to be transparent about the context of this experiment. As noted in A1, evaluating generative quality in the extreme data scarcity of FSCIL is challenging, and any single metric like FID should be interpreted with caution. We believe a deeper analysis of these two distinct paradigms for FSCIL is a promising direction for future research.

[a] "Calibrating higher-order statistics for few-shot class-incremental learning with pre-trained vision transformers." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.