FreeStyle: Free Control of Style-Content Dual-Reference Generation from Community LoRA Mining

Hugging Face Daily Papers Papers

Summary

FreeStyle proposes a scalable dual-reference generation framework using community LoRA mining to construct large-scale style-content triplets, with disentanglement mechanisms to prevent content leakage, and introduces a comprehensive benchmark for evaluation.

Style-content dual-reference generation aims to synthesize an image that preserves the structure and semantics of a content reference while adopting the style of a separate style reference.Despite recent progress, this setting remains challenging because models must balance content fidelity, style alignment, and instruction following avoiding semantic leakage from the style reference.A key bottleneck is the lack of large-scale triplet data with clean content-style separation and broad long-tail style coverage.In this work, we propose FreeStyle, a scalable dual-reference generation framework based on community LoRA mining.We treat community LoRAs as compositional anchors for style and content, and design a rigorous generation and filtering pipeline to construct large-scale Style-Reference and Content-Reference triplets across multiple base models.To address content leakage, we adopt a two-stage curriculum with stage-specific disentanglement mechanisms: an attention-level enrichment constraint that suppresses style-reference leakage in the style-transfer stage, and a frequency-aware RoPE modulation strategy that targets positional-correspondence-based leakage in the harder dual-reference stage.We also introduce a benchmark covering both style-reference and dual-reference generation, with evaluations on style similarity, content preservation, aesthetics, instruction following, and leakage rejection. The benchmark incorporates a style-invariant Content Alignment Score (CAS) and introduces a calibrated VLM-based Rejection Score for evaluating generation reliability and leakage suppression.Extensive experiments show that our model achieves a strong balance among style alignment, content preservation, and leakage suppression.
Original Article
View Cached Full Text

Cached at: 06/20/26, 02:29 PM

Paper page - FreeStyle: Free Control of Style-Content Dual-Reference Generation from Community LoRA Mining

Source: https://huggingface.co/papers/2606.20506 Authors:

,

,

,

,

,

,

,

,

,

,

Abstract

FreeStyle is a scalable dual-reference generation framework that uses community LoRA mining to create large-scale style-content triplets while addressing content leakage through disentanglement mechanisms and a comprehensive benchmark.

Style-contentdual-reference generationaims to synthesize an image that preserves the structure and semantics of a content reference while adopting the style of a separate style reference.Despite recent progress, this setting remains challenging because models must balance content fidelity, style alignment, and instruction following avoiding semantic leakage from the style reference.A key bottleneck is the lack of large-scale triplet data with clean content-style separation and broad long-tail style coverage.In this work, we propose FreeStyle, a scalabledual-reference generationframework based on communityLoRA mining.We treat community LoRAs as compositional anchors for style and content, and design a rigorous generation and filtering pipeline to construct large-scale Style-Reference and Content-Reference triplets across multiple base models.To addresscontent leakage, we adopt a two-stage curriculum with stage-specificdisentanglement mechanisms: anattention-level enrichment constraintthat suppresses style-reference leakage in the style-transfer stage, and afrequency-aware RoPE modulationstrategy that targets positional-correspondence-based leakage in the harder dual-reference stage.We also introduce a benchmark covering both style-reference anddual-reference generation, with evaluations on style similarity, content preservation, aesthetics, instruction following, and leakage rejection. The benchmark incorporates a style-invariantContent Alignment Score(CAS) and introduces a calibrated VLM-basedRejection Scorefor evaluating generation reliability and leakage suppression.Extensive experiments show that our model achieves a strong balance among style alignment, content preservation, and leakage suppression.

View arXiv pageView PDFProject pageGitHub13Add to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.20506 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.20506 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.20506 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Small Is Enough: Per-User Style Rewriting of AI-Edited Text via LoRA Adapters

arXiv cs.CL

This paper introduces InMyStyle, a privacy-first system that uses LoRA adapters on small language models (0.5B–7B) to rewrite AI-edited text toward an individual user's writing style without explicit prompts. Evaluations show quality plateaus across model sizes, suggesting compact models suffice for this task.

Interpreting Style Representations via Style-Eliciting Prompts

arXiv cs.CL

This paper proposes a framework to interpret style representations by using style-eliciting prompts—natural language instructions that steer LLMs to generate text with specific stylistic attributes. The method outperforms baseline LLM prompting techniques in both describing and imitating writing styles.

Style or Content? Evaluating Style Classifiers with Controlled Content Overlap

arXiv cs.CL

This paper introduces a controlled content overlap setup using parallel Bible translations to evaluate how much style classifiers rely on content cues rather than actual style features. Results show that low-overlap models degrade when content cues are removed, while high-overlap models transfer more robustly.