Research map / Video / Video generation

StyleMaster: Stylize Your Video with Artistic Generation and Translation

CVPR 2025 · Video › Video generation · approach: Diffusion-based human image animation · 16 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 StyleMaster, a DiT-based video generation model with local-patch and contrastive global style extraction (with a model-illusion paired dataset), a motion adapter used at negative scale, and a gray tile ControlNet for stylized generation and video style transfer.

Problem

Style-controlled video generation and video style transfer suffer from weak local texture preservation, content leakage from the reference image, and temporal/motion degradation when trained only on images.

Approach

Combine CLIP patch features filtered by low similarity to the prompt (texture) with a contrastive-trained MLP projection on CLIP image embeddings (global style), where contrastive pairs come from model-illusion pixel-rearranged images that share style but differ in content; add a still-video motion adapter whose negative scale boosts dynamics and stylization, plus a grayscale tile ControlNet for c

Key results

• Image style transfer: CSD-Score 0.45 vs 0.40 (StyleID) and ArtFID 36.89 vs 38.57 (StyleID), best on the first two metrics.
• Stylized video generation: CLIP-Text 0.305 and UMT-Score 2.329, outperforming StyleCrafter (0.294, 1.994) and VideoComposer; MotionSmooth 0.994.
• Global projection raises UMT score from 0.892 to 2.337 by reducing content leakage (ablation B1 vs B2).

Measured effects reported in the paper

Surprising findings

• VideoComposer achieves a higher CSD score (0.680 vs 0.463) because it directly copies reference content, which the authors argue inflates style scores.
• Negative motion adapter scale both increases style similarity and dynamics, but harms text alignment and motion smoothness beyond -0.3.

Limitations

• Relies on reference style images; does not extract or transfer dynamic styles such as particle effects or motion from reference videos.
• Slightly underperforms baselines on content-preservation metrics (LPIPS, CFSD) in image style transfer.
• Negative motion adapter scale beyond -0.3 degrades text alignment and motion smoothness.

Experimental setup

Datasets: Laion-Aesthetics model-illusion style pairs (self-generated) still videos StyleCrafter test set (expanded)

Benchmarks: StyleCrafter test set (expanded to 192 style-prompt pairs) image style transfer test set (96 content-style pairs)

Metrics: CSD-Score ArtFID FID LPIPS CFSD CLIP-Text UMT-Score VisualQuality DynamicQuality MotionSmooth

Baselines: VideoComposer StyleCrafter InstantStyle CSGO StyleID DomoAI AnyV2V

Builds on: AnimateDiff StillMoving StyleCrafter IP-Adapter Visual Anagrams model illusion

Closest papers by meaning

• CSGO: Content-Style Composition in Text-to-Image Generation (NeurIPS 2025 · similarity 0.87)
• Conditional Balance: Improving Multi-Conditioning Trade-Offs in Image Generation (CVPR 2025 · similarity 0.87)
• Semantix: An Energy-guided Sampler for Semantic Style Transfer (ICLR 2025 · similarity 0.87)
• OmniStyle: Filtering High Quality Style Transfer Data at Scale (CVPR 2025 · similarity 0.86)
• SigStyle: Signature Style Transfer via Personalized Text-to-Image Models (AAAI 2025 · similarity 0.86)
• CLeaR: A Unified Framework for Resolving the Leakage–Degradation Dilemma in Style Transfer (NeurIPS 2026 · similarity 0.86)