Research map / Video / Video generation

ByTheWay: Boost Your Text-to-Video Generation Model to Higher Quality in a Training-free Way

CVPR 2025 · Video › Video generation · approach: Compositional text-to-video diffusion · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ByTheWay, a training-free inference-time method for text-to-video diffusion models that reduces inter-block temporal attention disparity (Temporal Self-Guidance) and amplifies high-frequency temporal attention energy (Fourier-based Motion Enhancement) to improve structure, consistency and motion.

Problem

Text-to-video diffusion models often produce structurally implausible, temporally inconsistent, or nearly static videos, and prior fixes require costly retraining or add inference cost.

Approach

Observes that inter-block disparity of temporal attention maps correlates with artifacts and that attention-map energy correlates with motion magnitude; mitigates disparity by injecting the up_blocks.1 temporal attention map into later blocks, and boosts motion by scaling high-frequency FFT components of the softmax-normalized attention map, with no extra parameters or sampling cost.

Key results

• On VBench, AnimateDiff + ByTheWay raises Dynamic Degree from 0.4073 to 0.5245 and Motion Smoothness from 0.9474 to 0.9786.
• VideoCrafter2 + ByTheWay raises Dynamic Degree from 0.4086 to 0.5547 and Subject Consistency from 0.9732 to 0.9852.
• In a 30-person user study, ByTheWay received 74.58% of votes for AnimateDiff video quality versus 25.42% for vanilla.

Measured effects reported in the paper

Surprising findings

• FreeInit, a competing training-free method, lowers AnimateDiff Dynamic Degree from 0.4073 to 0.2941 while improving consistency metrics.

Limitations

• Default hyperparameters (alpha, beta, tau) must be tuned per backbone and are not universally optimal.
• Excessively large alpha or beta can cause loss of motion or motion artifacts.

Experimental setup

Datasets: VBench prompts

Benchmarks: VBench

Metrics: user study vote share GPT-4o MLLM assessment vote share VBench subject consistency background consistency motion smoothness dynamic degree aesthetic quality imaging quality

Baselines: AnimateDiff VideoCrafter2 FreeInit UniCtrl I4VGen FreeU

Builds on: AnimateDiff VideoCrafter2 FreeInit FreeU latent diffusion model

Closest papers by meaning

• StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.90)
• Enhancing Motion in Text-to-Video Generation with Decomposed Encoding and Conditioning (NeurIPS 2024 · similarity 0.89)
• A Recipe for Scaling up Text-to-Video Generation with Text-free Videos (CVPR 2024 · similarity 0.89)
• LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity (CVPR 2025 · similarity 0.89)
• Encapsulated Composition of Text-to-Image and Text-to-Video Models for High-Quality Video Synthesis (CVPR 2025 · similarity 0.89)
• EIDT-V: Exploiting Intersections in Diffusion Trajectories for Model-Agnostic, Zero-Shot, Training-Free Text-to-Video Generation (CVPR 2025 · similarity 0.89)