pith. sign in

arxiv: 2606.03183 · v1 · pith:3VJBIKB7new · submitted 2026-06-02 · 💻 cs.MM · cs.CV· cs.SD· eess.AS

Inference-Time Scaling for Joint Audio-Video Generation

Jaemin Jung , Kyeongha Rho , Inkyu Shin , Joon Son Chung This is my paper

Pith reviewed 2026-06-28 07:47 UTC · model grok-4.3

classification 💻 cs.MM cs.CVcs.SDeess.AS
keywords inference-time scalingjoint audio-video generationmulti-verifier frameworkadaptive reward weightingsemantic alignmentaudio-visual synchronizationtest-time optimization
0
0 comments X

The pith

A multi-verifier framework plus adaptive reward weighting enables effective inference-time scaling for joint audio-video generation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper investigates how to apply inference-time scaling to the joint generation of audio and video from text prompts. It establishes that single-objective guidance creates uneven trade-offs and allows models to exploit individual verifiers, so multiple verifiers must be combined. The authors then introduce Adaptive Reward Weighting, an online optimization step that learns to balance the different reward signals on the fly. Experiments on VGGSound and JavisBench-mini show gains in semantic alignment, perceptual quality, and audio-visual timing. This matters because it offers a training-free route to higher-quality multimodal outputs.

Core claim

The authors claim that inference-time scaling for joint audio-video generation requires a multi-verifier setup to avoid asymmetric trade-offs and verifier hacking, and that Adaptive Reward Weighting aggregates the resulting signals by treating reward balancing as an online optimization problem with learnable parameters that calibrate variances without prior distributional knowledge.

What carries the argument

Adaptive Reward Weighting (ARW), a test-time algorithm that performs online optimization over learnable parameters to calibrate and combine multiple reward signals during generation.

If this is right

  • The chosen multi-verifier combination produces balanced gains across semantic alignment, perceptual quality, and audio-visual synchronization.
  • Adaptive Reward Weighting aggregates heterogeneous rewards without needing advance knowledge of their distributions.
  • The resulting outputs improve on VGGSound and JavisBench-mini relative to prior joint generation baselines.
  • All improvements occur at inference time without additional model training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same multi-verifier structure could be tested on other pairs of modalities whose quality metrics conflict during generation.
  • If ARW's learnable parameters converge reliably, the method may reduce the need for hand-tuned reward weights in future multimodal systems.
  • The approach implies that test-time compute can substitute for some of the training compute currently spent on multimodal alignment.

Load-bearing premise

Single-objective guidance necessarily produces asymmetric performance trade-offs and verifier hacking, so multiple verifiers are required.

What would settle it

An experiment that applies single-verifier inference-time scaling to the same base model and benchmarks and measures equal or superior scores across semantic alignment, perceptual quality, and synchronization would falsify the necessity of the multi-verifier approach.

Figures

Figures reproduced from arXiv: 2606.03183 by Inkyu Shin, Jaemin Jung, Joon Son Chung, Kyeongha Rho.

Figure 1
Figure 1. Figure 1: Qualitative results with Inference-Time Scaling. Compared to naive sampling, ITS yields audio–video pairs with superior semantic alignment with the text prompt and cross-modal synchronization. The video samples are available at the following link. Joint audio-video generation aims to simultaneously synthesize auditory and visual streams that are tem￾porally synchronized and semantically aligned. MM-Diffusi… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of reward guidance types and aggregation methods. The overall performance improvement is calculated by averaging the relative improvements of all evaluation metrics compared to the naive sampling (without ITS). This result is obtained with JavisDiT. where {βt} T t=1 denotes the noise schedule and ϵ represents standard Gaussian noise. Then, a conditional denoising network ϵθ(xt, s, t) is trained … view at source ↗
Figure 3
Figure 3. Figure 3: Overview of Adaptive Reward Weighting. Given audio-video samples generated by the pre￾trained generative model, they are evaluated by multiple verifiers, with their raw rewards stored in a history buffer. Then, the aggregated score is computed via a weighted sum of these rewards using reward-specific learnable calibration parameters, and calibration parameters are updated through test-time optimization. Th… view at source ↗
Figure 4
Figure 4. Figure 4: Inference-time scaling curves across the number of samples. Across both Best-of-N and EvoSearch, the Z-score and ARW aggregation methods exhibit consistent performance improvement as the sample size increases. In both (a) and (b), the left and right panels show VR and JS scores, respectively. This balance is particularly critical for an iterative search algorithm. The performance gap between Best￾of-N and … view at source ↗
Figure 5
Figure 5. Figure 5: Human evaluation results. Impact of Inference-Time Scaling. To evaluate the scalability of the proposed methods, we apply Z￾score normalization and ARW to two ITS strategies while gradually increasing the samples. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of generated samples. We compare the outputs of naive sampling, single-verifier (VR-guidance), and our multi-verifier (ARW) given a complex text prompt on JavisDiT. The video samples are available at the following link. samples. These subjective findings further corroborate the objective evaluations reported in [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Sensitivity analysis of preference weights. We vary the preference weights in Eq. 3 from 0.3 to 0.7 in increments of 0.1. Increasing the weight for VideoReward-TA improves text consistency, whereas emphasizing JavisScore enhances AV consistency. Notably, the method demonstrates robustness, maintaining stable performance in both modalities without sudden degradation across the tested range. D.3 Hyperparamet… view at source ↗
Figure 8
Figure 8. Figure 8: Convergence behavior of ARW under different optimizers. We plot the calibration loss (top) and the learned scale parameters (bottom) for Adam, SGD, and RMSprop. All optimizers converge to similar final solutions, although their early optimization dynamics differ. Overall, the results indicate that ARW is stable and largely insensitive to optimizer choice. D.5 Generalization to a Stronger AV-Generation Mode… view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison of generated samples. We compare the outputs of naive sampling, single-verifier (VR-guidance), and our multi-verifier (ARW) given a complex text prompt on JavisDiT. Naive sampling fails to respect the count constraint, generating three pigeons. Single-verifier guidance misses the fine-grained visual detail, failing to render the mesh fence. Multi-verifier successfully satisfies all c… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison of generated samples. We compare the outputs of naive sampling, single-verifier (VR-guidance), and our multi-verifier (ARW) given a complex text prompt on JavisDiT. Naive sampling fails to generate the specific audio event. Single-verifier misinterprets the audio description as a visual cue, hallucinating a bird’s wing in the video frame. Multi-verifier correctly assigns the semanti… view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison of generated samples. We compare the outputs of naive sampling and our multi-verifier (ARW) approach using text prompts from MMDisCo. The results demonstrate the correction of semantic and physical failures. In the top example, naive sampling generates a bird with severe anatomical distortion (head twisted backwards), whereas our method ensures physical correctness and generates syn… view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative results on prompt variability. We compare the videos conditioned on the original JavisBench-mini prompt and its three transformed variants: conversational style, fragment-style, and grammar-typo prompts. Without ITS, naive sampling often fails to preserve key prompt details, such as the visible pouring stream and the surface dynamics of the oil. In contrast, ITS generates videos that more cons… view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative results on prompt variability. We compare the outputs conditioned on the original JavisBench-mini prompt and its three transformed variants: conversational style, fragment-style, and grammar-typo prompts. Without ITS, naive sampling often fails to preserve key prompt semantics under prompt variation: it loses the intended interaction between the two birds, hallucinates an additional orange bir… view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparison on LTX-2. We compare naive sampling and multi-verifier ITS with ARW on a stronger open-source joint audio-video generation model. Top: ITS better satisfies the fine￾grained bird attributes and count. Bottom: ITS better follows the intended helicopter motion direction. These examples suggest that ARW improves prompt-faithful audiovisual generation on LTX-2 as well. The video samples … view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative comparison on LTX-2. We compare naive sampling and multi-verifier ITS with ARW on a stronger open-source joint audio-video generation model. Top: ITS better reflects the intended piano-playing posture and interaction with the instrument. Bottom: ITS produces more distinct footstep events that better align with the walking motion. These examples suggest that ARW improves prompt￾faithful audiovi… view at source ↗
Figure 16
Figure 16. Figure 16: Failure cases. Top: Although the ITS framework generally improves the visual quality and text alignment, it still struggles with physical plausibility. For instance, the generated bird’s wing unnaturally passes through the solid feeder, indicating a lack of physical priors. Bottom: ITS successfully corrects semantic errors, such as guiding the horse to properly walk forward, unlike the naive sampling whic… view at source ↗
read the original abstract

Joint audio-video generation aims to synthesize realistic audio-video pairs that are both semantically aligned with text prompts and precisely synchronized. While existing joint audio-video generation models often require substantial training resources to improve fidelity, Inference-Time Scaling (ITS) has recently emerged as a promising training-free alternative in single-modality domains. However, extending ITS from a single modality to multimodal domains is non-trivial, as it requires balancing multiple heterogeneous objectives. In this paper, we present the first comprehensive study of ITS for joint audio-video generation. We first demonstrate that a multi-verifier framework is essential to address the limitations of single-objective guidance, including asymmetric performance trade-offs and verifier hacking. Through systematic analysis, we then identify an optimal multi-verifier combination that yields balanced improvements across all quality dimensions. Finally, to effectively aggregate diverse reward signals, we propose Adaptive Reward Weighting (ARW), a novel test-time optimization algorithm. ARW treats reward aggregation as an online optimization problem, utilizing learnable parameters to calibrate reward variances without requiring prior knowledge of reward distributions, thereby ensuring robust multi-objective selection. Experimental results on VGGSound and JavisBench-mini benchmarks demonstrate that our framework significantly enhances semantic alignment, perceptual quality, and audio-visual synchronization of generated outputs. Synthesized samples and code are available on the project page: https://jung-jaemin.github.io/ITS-AVGen-Proj.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper conducts the first comprehensive study of Inference-Time Scaling (ITS) for joint audio-video generation. It argues that single-objective guidance leads to asymmetric performance trade-offs and verifier hacking, necessitating a multi-verifier framework. An optimal verifier combination is identified through systematic analysis, and Adaptive Reward Weighting (ARW) is introduced as a test-time optimization algorithm that treats reward aggregation as an online problem with learnable parameters to calibrate variances without prior distribution knowledge. Experiments on VGGSound and JavisBench-mini are claimed to show significant gains in semantic alignment, perceptual quality, and audio-visual synchronization.

Significance. If substantiated, this would represent a notable advance by extending training-free ITS methods to the multimodal AV domain, where balancing heterogeneous objectives is challenging. The ARW approach is a distinct contribution for distribution-agnostic reward aggregation via online learning. The release of code and samples supports reproducibility and allows direct verification of the framework.

major comments (2)
  1. [Abstract] Abstract: The central claim that 'our framework significantly enhances semantic alignment, perceptual quality, and audio-visual synchronization of generated outputs' is asserted without any quantitative results, metric values, baseline comparisons, ablation details, or error analysis. This directly undermines evaluation of whether the data support the empirical conclusions.
  2. [Abstract] Abstract: The assertion that 'a multi-verifier framework is essential' due to asymmetric trade-offs and verifier hacking is presented as demonstrated, but the systematic analysis, specific evidence, or data showing these limitations (and the optimality of the chosen combination) are not provided in the text.
minor comments (1)
  1. [Abstract] Abstract: The project page is referenced but without enumeration of available assets (e.g., specific generated samples, code release details, or benchmark subsets).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed feedback on the abstract. We agree that the abstract can be strengthened by incorporating key quantitative highlights and brief references to supporting analysis from the body of the paper. We will revise the abstract accordingly while preserving its summary nature. Below we respond to each major comment.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that 'our framework significantly enhances semantic alignment, perceptual quality, and audio-visual synchronization of generated outputs' is asserted without any quantitative results, metric values, baseline comparisons, ablation details, or error analysis. This directly undermines evaluation of whether the data support the empirical conclusions.

    Authors: We acknowledge that the abstract presents high-level claims without specific numbers. The full manuscript provides these details in Section 4 (Experiments), including tables with metric values (e.g., improvements in semantic alignment, perceptual quality, and AV synchronization on VGGSound and JavisBench-mini), baseline comparisons, and ablations. To address the concern, we will revise the abstract to include concise quantitative highlights from the main results. revision_made: yes revision: yes

  2. Referee: [Abstract] Abstract: The assertion that 'a multi-verifier framework is essential' due to asymmetric trade-offs and verifier hacking is presented as demonstrated, but the systematic analysis, specific evidence, or data showing these limitations (and the optimality of the chosen combination) are not provided in the text.

    Authors: The abstract summarizes our finding that a multi-verifier framework is essential. The systematic analysis demonstrating asymmetric trade-offs, verifier hacking, and the optimal combination is presented in Section 3, supported by figures and tables. We will revise the abstract to briefly note the key evidence from this analysis (e.g., reference to observed trade-offs). revision_made: yes revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical framework validated by benchmarks

full rationale

The paper presents an empirical study of inference-time scaling for joint audio-video generation. It demonstrates limitations of single-objective guidance via analysis, identifies an optimal multi-verifier combination, and proposes ARW as a test-time optimization method using learnable parameters. All central claims (improved semantic alignment, perceptual quality, and synchronization) are supported by experimental results on VGGSound and JavisBench-mini benchmarks rather than any derivation chain. No equations, fitted predictions, self-citations, or ansatzes are described that reduce the outputs to the paper's own inputs by construction. The argument is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 1 invented entities

Review performed on abstract only; full implementation details unavailable.

free parameters (1)
  • learnable parameters in ARW
    Used to calibrate reward variances during online optimization without prior knowledge of reward distributions
invented entities (1)
  • Adaptive Reward Weighting (ARW) no independent evidence
    purpose: Aggregate diverse reward signals from multiple verifiers as an online optimization problem
    Newly introduced test-time algorithm

pith-pipeline@v0.9.1-grok · 5787 in / 998 out tokens · 21237 ms · 2026-06-28T07:47:44.410494+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 13 canonical work pages · 7 internal anchors

  1. [1]

    Large Language Monkeys: Scaling Inference Compute with Repeated Sampling

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christopher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling.arXiv preprint arXiv:2407.21787,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    LTX-2: Efficient Joint Audio-Visual Foundation Model

    Yoav HaCohen, Benny Brazowski, Nisan Chiprut, Yaki Bitterman, Andrew Kvochko, Avishai Berkowitz, Daniel Shalem, Daphna Lifschitz, Dudu Moshe, Eitan Porat, et al. Ltx-2: Efficient joint audio-visual foundation model.arXiv preprint arXiv:2601.03233,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Scaling image and video generation via test-time evolutionary search.arXiv preprint arXiv:2505.17618, 2025a

    Haoran He, Jiajun Liang, Xintao Wang, Pengfei Wan, Di Zhang, Kun Gai, and Ling Pan. Scaling image and video generation via test-time evolutionary search.arXiv preprint arXiv:2505.17618, 2025a. Xuan He, Dongfu Jiang, Ping Nie, Minghao Liu, Zhengxuan Jiang, Mingyi Su, Wentao Ma, Junru Lin, Chun Ye, Yi Lu, et al. Videoscore2: Think before you score in genera...

    work page arXiv

  4. [4]

    Imagen Video: High Definition Video Generation with Diffusion Models

    Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303,

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    Inference-time scaling for diffusion-based audio super-resolution.arXiv preprint arXiv:2508.02391,

    Yizhu Jin, Zhen Ye, Zeyue Tian, Haohe Liu, Qiuqiang Kong, Yike Guo, and Wei Xue. Inference-time scaling for diffusion-based audio super-resolution.arXiv preprint arXiv:2508.02391,

    work page arXiv

  6. [6]

    Voicedit: Dual-condition diffusion transformer for environment-aware speech synthesis

    Jaemin Jung, Junseok Ahn, Chaeyoung Jung, Tan Dat Nguyen, Youngjoon Jang, and Joon Son Chung. Voicedit: Dual-condition diffusion transformer for environment-aware speech synthesis. InProc. ICASSP, 2025a. Jaemin Jung, Jaehun Kim, Inkyu Shin, and Joon Son Chung. Score: Scaling audio generation using stan- dardized composite rewards.arXiv preprint arXiv:2509...

    work page arXiv 2001

  7. [7]

    Adam: A Method for Stochastic Optimization

    Jaihoon Kim, Taehoon Yoon, Jisung Hwang, and Minhyuk Sung. Inference-time scaling for flow models via stochastic generation and rollover budget forcing. InProc. NeurIPS, 2025a. Sunwoo Kim, Minkyu Kim, and Dongmin Park. Test-time alignment of diffusion models without reward over-optimization. InProc. ICLR, 2025b. Diederik P Kingma and Jimmy Ba. Adam: A met...

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Syncflow: Toward temporally aligned joint audio-video generation from text.arXiv preprint arXiv:2412.15220,

    Haohe Liu, Gael Le Lan, Xinhao Mei, Zhaoheng Ni, Anurag Kumar, Varun Nagaraja, Wenwu Wang, Mark D Plumbley, Yangyang Shi, and Vikas Chandra. Syncflow: Toward temporally aligned joint audio-video generation from text.arXiv preprint arXiv:2412.15220,

    work page arXiv

  9. [9]

    Improving Video Generation with Human Feedback

    Jie Liu, Gongye Liu, Jiajun Liang, Ziyang Yuan, Xiaokun Liu, Mingwu Zheng, Xiele Wu, Qiulin Wang, Wenyu Qin, Menghan Xia, et al. Improving video generation with human feedback.arXiv preprint arXiv:2501.13918, 2025b. Kai Liu, Wei Li, Lai Chen, Shengqiong Wu, Yanhao Zheng, Jiayi Ji, Fan Zhou, Rongxin Jiang, Jiebo Luo, Hao Fei, et al. Javisdit: Joint audio-v...

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Ovi: Twin Backbone Cross-Modal Fusion for Audio-Video Generation

    Yexiang Liu, Zekun Li, Zhi Fang, Nan Xu, Ran He, and Tieniu Tan. Rethinking the role of prompting strategies in llm test-time scaling: A perspective of probability theory. InProc. ACL, 2025c. Chetwin Low, Weimin Wang, and Calder Katyal. Ovi: Twin backbone cross-modal fusion for audio-video generation.arXiv preprint arXiv:2510.01284,

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Movie Gen: A Cast of Media Foundation Models

    Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-YaoMa, Ching-YaoChuang, etal. Moviegen: Acastofmediafoundationmodels.arXiv preprint arXiv:2410.13720,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Av-dit: Efficient audio-visual diffusion transformer for joint audio and video generation.arXiv preprint arXiv:2406.07686, 2024a

    Kai Wang, Shijian Deng, Jing Shi, Dimitrios Hatzinakos, and Yapeng Tian. Av-dit: Efficient audio-visual diffusion transformer for joint audio and video generation.arXiv preprint arXiv:2406.07686, 2024a. Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-la...

    work page arXiv

  13. [13]

    Uniform: A unified multi-task diffusion transformer for audio-video generation.arXiv preprint arXiv:2502.03897,

    16 Lei Zhao, Linfeng Feng, Dongxu Ge, Rujin Chen, Fangqiu Yi, Chi Zhang, Xiao-Lei Zhang, and Xuelong Li. Uniform: A unified multi-task diffusion transformer for audio-video generation.arXiv preprint arXiv:2502.03897,

    work page arXiv

  14. [14]

    Transfusion: Predict the next token and diffuse images with one multi-modal model

    Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. InProc. ICLR, 2025a. Zikai Zhou, Shitong Shao, Lichen Bai, Shufei Zhang, Zhiqiang Xu, Bo Han, and Zeke Xie. Golden noise for diffus...

    2023

  15. [15]

    Despite these advances in training strategies, the ability to adaptively refine generation at inference time without costly retraining remains underexplored

    further prioritizes fine-grained spatio-temporal synchronization by injecting hierarchical priors directly into DiT blocks. Despite these advances in training strategies, the ability to adaptively refine generation at inference time without costly retraining remains underexplored. Inference-Time Scaling for Diffusion Models.Inspired by the success of Larg...

    2025

  16. [16]

    For instance, the latent beam search (Oshima et al.,

    or iteratively optimizing latent frequency components (Wu et al., 2024; Yuan et al., 2025); and (ii) search-based selection, which explores multiple candidates to identify the optimal output according to a specific scoring function. For instance, the latent beam search (Oshima et al.,

    2024

  17. [17]

    maintains a set of promising latent candidates at each denoising step and prunes lower-quality paths to concentrate compute on high-reward trajectories. On the other hand, the evolutionary search (He et al., 2025a) reformulates the sampling process as an evolution- ary optimization problem, applying selection and mutation mechanisms to intermediate denois...

    2018

  18. [18]

    Despite their effectiveness, these approaches require access to the model’s internal parameters and involve computationally heavy gradient calculations over a training dataset

    dynamically weights loss functions during training based on task-specific uncertainty. Despite their effectiveness, these approaches require access to the model’s internal parameters and involve computationally heavy gradient calculations over a training dataset. In contrast, our proposed ARW introduces an inference-time paradigm that does not require upd...

    2018

  19. [19]

    We compare Adam (Kingma & Ba, 2014), SGD (Robbins & Monro, 1951), and RMSprop (Tieleman,

    2014

  20. [20]

    two pigeons

    by tracking both the calibration loss (Larw) and the learned scale parameters over optimization steps. All three optimizers converge to very similar final loss values and nearly identical scale parameters, indicating that ARW is insensitive to the choice of optimizer. In particular, the learned scales stabilize within roughly 50–100 steps in all cases, su...

    2026

  21. [21]

    First,physical plausibility is still not guaranteed. In the upper example, ITS improves the overall prompt alignment by generating a hummingbird feeding from the feeder, but the bird’s wings still unrealistically pass through the feeder, indicating a violation of basic physics. Second,unstable temporal consistency remains an issue. In the lower example, n...

    2026