Source-linked AI summary
NoisEasier: Test-Time Noise Optimization for Text-to-Video Generation
Yujiang Pu, Yu Kong
TL;DR
Text-to-video models remain unreliable on fine-grained compositional alignment, while reward-based fine-tuning is difficult to adapt and vulnerable to reward hacking. NoisEasier performs differentiable reward-guided noise optimization at test time using efficient generators and robust rewards, with joint trajectory optimization improving alignment and complementing offline fine-tuning. The method improves compositional alignment while preserving visual fidelity and adds negligible computational cost, although performance remains bounded by reward-model quality.
Problem
Text-to-video models struggle with fine-grained compositional alignment, while reward-based fine-tuning is expensive to adapt and susceptible to reward hacking.
Method
NoisEasier performs test-time differentiable reward-guided noise optimization using short-step Video Consistency Models and robust multi-objective rewards, without model fine-tuning.
Results
NoisEasier consistently improves compositional alignment while preserving visual fidelity, and jointly optimizing intermediate noises outperforms initial-latent optimization with negligible additional computational cost.
Takeaways & Limitations
NoisEasier functions as both an alternative and a complement to offline reward fine-tuning, retaining effectiveness after model alignment.
Takeaways & Limitations
Test-time optimization is bounded by reward-model quality, with current VLM-based rewards remaining insufficiently sensitive to exact object counting and complex compositional reasoning.
Abstract
from arXiv · showhide
Diffusion models have recently advanced text-to-video (T2V) generation, yet they still struggle with fine-grained compositional alignment, such as attribute binding, spatial relations, and object interactions. While reward-based fine-tuning improves alignment, it is susceptible to reward hacking and adapts poorly to new prompt distributions. In this work, we propose NoisEasier, a test-time scaling framework that improves T2V generation through differentiable reward-guided noise optimization without modifying the underlying model. By combining efficient short-step generators with a multi-objective reward formulation, NoisEasier enables stable and practical test-time optimization under realistic inference budgets. Our key insight is that jointly optimizing the entire stochastic trajectory accelerates reward convergence and improves compositional alignment over optimizing only the initial latent, with negligible additional computational and time cost. Experiments on VBench and T2V-CompBench demonstrate consistent improvements across multiple backbones, achieving over 10% average gains on challenging dimensions such as attribute binding, object interaction, and numeracy. Overall, NoisEasier serves as both a flexible alternative and a complementary enhancement to reward-based fine-tuning, establishing test-time scaling as an effective paradigm for controllable text-to-video generation.
1 Introduction
NoisEasier addresses fine-grained compositional alignment challenges in text-to-video by optimizing noise at test time without model fine-tuning. It combines short-step generation, robust reward supervision, and joint optimization of stochastic perturbations to improve alignment efficiently.
- Motivation: Text-to-video systems still struggle with attribute binding, object interactions, and motion dynamics while preserving visual fidelity.
- Motivation: Reward-based fine-tuning improves global alignment but is costly to adapt and remains vulnerable to reward hacking.
- Motivation: Long denoising chains, high memory use, and fragile video rewards make iterative gradient-based test-time refinement difficult for T2V models.
- Approach: NoisEasier uses short-step Video Consistency Models and a robust multi-objective reward formulation for practical, stable test-time optimization.
- Approach: Jointly optimizing intermediate stochastic perturbations accelerates reward convergence and strengthens compositional alignment over optimizing only initial noise, with negligible extra cost.
- Results: Experiments on VBench and T2V-CompBench show improvements across attribute binding, spatial relations, and object interactions, including gains beyond reward-finetuned models.
2 Related Work
Prior text-to-video alignment work largely follows preference-driven optimization or reward-based approaches. These methods improve alignment but require substantial preference data and have limited flexibility for incorporating structured human knowledge.
- Human-feedback approaches to video generation draw on RLHF, extending alignment methods developed for large language models.
- Preference-driven optimization trains models using large-scale human preference datasets, including policy-gradient and direct-preference methods.
- Preference-based approaches require extensive annotations and cannot easily incorporate structured human knowledge.
3 Method
NoisEasier makes test-time T2V refinement practical by combining short-step consistency models with differentiable, multi-objective reward optimization. It jointly optimizes initial noise and intermediate stochastic perturbations, while using video, motion, and negative-aware rewards for finer compositional control.
- Efficient test-time optimization: Video Consistency Models reduce gradient-based refinement costs by distilling long diffusion sampling into only a few denoising steps.This enables practical end-to-end optimization for video generation.
- Joint stochastic optimization: Jointly optimizing initial latent noise and intermediate stochastic perturbations provides finer denoising-trajectory control and faster reward convergence than optimizing initialization alone.The method reports negligible additional computational overhead.
- Multi-objective rewards: The reward formulation combines image-level semantic, video-text, and motion-aware signals to improve semantic alignment while preserving localized, temporally coherent dynamics.The motion reward suppresses global camera motion and penalizes abrupt changes in flow magnitude.
- Negative-aware calibration: Negative-aware calibration contrasts the target prompt with syntactically similar but semantically incorrect negatives to make rewards sensitive to compositional errors.The calibrated objective uses a margin and temperature to penalize semantically similar distractors.
- Optimization objective: The method optimizes all stochastic variables through gradient ascent on a weighted combination of calibrated reward components.The optimization variables include zT and the intermediate perturbations ε1 through εT.
4 Experiment
Experiments across VBench and T2V-CompBench show that NoisEasier improves compositional alignment across multiple VCM backbones, while reward combinations and full-trajectory optimization provide complementary benefits. The method adds negligible overhead, complements reward fine-tuning, and remains limited by reward-model sensitivity to temporal transitions and exact counting.
- Main Results: NoisEasier improves VBench performance across AnimateLCM and T2V-Turbo, including Multiple Objects, Spatial Relation, and Dynamic Degree.On AnimateLCM, Overall Consistency rises from 25.74% to 28.79%; Multiple Objects rises from 54.50% to 76.62% on T2V-Turbo (MS) and from 54.33% to 79.18% on T2V-Turbo (VC2).
- Main Results: NoisEasier yields consistent T2V-CompBench gains in attribute binding, spatial relationships, interactions, and numeracy.Consistent Attribute Binding reaches 0.8375 on MS and 0.8737 on VC2, while Numeracy gains 0.0953 on MS.
- Reward Analysis: Semantic and motion rewards emphasize different outcomes, so the multi-objective formulation balances semantic alignment with temporal dynamics.Semantic rewards can produce overly static videos, whereas motion rewards increase dynamics but degrade semantic metrics.
- Optimization Analysis: Full trajectory optimization accelerates reward convergence and improves fine-grained structure over initial-noise optimization with negligible extra cost.Initial noise mainly controls global appearance and coarse layout, while intermediate perturbations refine structures and stylistic details.
- Comparison and Efficiency: NoisEasier complements reward fine-tuning, improving compositional metrics beyond offline tuning while incurring less than 1% additional runtime for full trajectory optimization.Reward fine-tuning mainly improves motion metrics, whereas NoisEasier provides larger gains on Multiple Objects and Color; applying it to the fine-tuned model improves most metrics further.
- Limitations: The method improves qualitative semantic, spatial, and temporal coherence but still fails on gradual temporal state transitions and precise object counting.These limitations are attributed to reward models that provide weak supervision for long-term evolution and insufficient sensitivity to exact counting.
5 Conclusion
NoisEasier improves text-to-video compositional alignment through differentiable reward-guided noise optimization without model fine-tuning. Jointly optimizing intermediate noises strengthens control over the denoising trajectory, while test-time optimization complements offline reward fine-tuning.
- NoisEasier improves text-to-video generation through differentiable reward-guided noise optimization without modifying model parameters.
- Jointly optimizing intermediate noises provides faster reward convergence and stronger compositional alignment than optimizing only the initial latent.
- NoisEasier preserves visual fidelity while consistently enhancing compositional alignment.
- Test-time optimization remains effective after offline reward fine-tuning, making NoisEasier complementary to model alignment.
- NoisEasier establishes noise optimization as a practical test-time scaling paradigm for improving alignment in video generative models.
A More Implementation Details
The implementation details calibrate heterogeneous reward models and construct hard negatives for reward supervision. Reward weights account for differing score scales, while automated prompt variation changes one semantic category at a time.
- Reward models receive separate weights because their output scales differ across semantic, aesthetic, and motion signals.HPSv2 and ViCLIP scores typically range from 0.2 to 0.4 and receive weight 2; ImageReward is min-max normalized from [−2, +2] to [0, 1] and receives weight 1.
- The hard-negative pipeline uses GPT-4o to generate syntactically similar but semantically different prompt variants.Each negative modifies only one category, such as the subject, object, or action.
B More Ablation Studies
Reward-model ablations show complementary strengths and trade-offs across semantic alignment, appearance, spatial fidelity, and motion. Combining motion-aware rewards with semantic or visual-text rewards provides a strong balance, while some combinations suppress motion expressiveness.
- ViCLIP delivers the strongest gains in overall consistency and color fidelity among single-reward settings.
- ImageReward performs best on multiple objects and spatial relations, reflecting its fine-grained visual-text training.
- The motion reward substantially boosts dynamic degree and smoothness despite trailing other rewards on appearance metrics.
- Pairing the motion reward with ViCLIP or ImageReward balances temporal dynamics with spatial fidelity.
- HPSv2 combined with other rewards can improve color quality while suppressing motion expressiveness.
- The selected configuration achieves competitive semantic alignment and strong object- and motion-related scores with modest variation across settings.
C Additional Results on Modern Video Generators
Additional experiments evaluate NoisEasier on modern distilled video generators under practical computational constraints. The method improves both FastWan and LTX-Video-2B-distilled, suggesting generalization beyond the VCM backbones used in the main paper.
- Applying NoisEasier directly to large diffusion backbones with long denoising chains remains computationally prohibitive.A single 5-second Wan2.1-14B video requires approximately 56 minutes and 69 GB peak VRAM on one A100 GPU.
- NoisEasier consistently improves both FastWan and LTX-Video-2B-distilled under the default optimization setting.
- 2.22% improvement in Overall Consistency is reported for FastWan, alongside gains in object recognition and color consistency.
- LTX-Video-2B-distilled gains in overall consistency, action understanding, and object interaction.
- Qualitative examples show improved semantic alignment across diverse prompts, supporting effectiveness beyond the VCM backbones studied in the main paper.
D More Visualization Results
The paper provides additional qualitative examples to illustrate NoisEasier’s effectiveness, spanning comparisons on VBench and T2V-CompBench.
- Additional qualitative examples are presented to illustrate the effectiveness of NoisEasier.These examples extend the visualization results beyond the earlier comparisons.
E Limitations and Future Work
NoisEasier adds inference overhead, remains constrained by the underlying generator and reward proxies, and has been evaluated primarily on short, single-scene videos. Future work therefore targets stronger rewards and more efficient optimization for longer-horizon generation.
- Limitations: Additional inference overhead makes NoisEasier less suitable for latency-sensitive applications.The limitation arises from its test-time optimization relative to vanilla sampling.
- Limitations: NoisEasier primarily improves semantic alignment, while visual quality remains bounded by the backbone and imperfect reward proxies.
- Future Work: The study focuses on short, single-scene videos, leaving longer videos, complex narratives, and higher resolutions challenging.The stated challenges include weaker reward signals and increased optimization cost.
- Future Work: Future work will explore stronger video-level rewards and more efficient optimization strategies for long-horizon video generation.