Hugging Face Daily Papers · June 19, 2026 · 4 min read

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

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

📚 Resources 📄 Paper: <a href=\"https://arxiv.org/pdf/2606.20506\" rel=\"nofollow\">https://arxiv.org/pdf/2606.20506</a> 🌐 Project Page: <a href=\"https://blue2giant.github.io/FreeStyle/\" rel=\"nofollow\">https://blue2giant.github.io/FreeStyle/</a> 💻 GitHub: <a href=\"https://github.com/Blue2Giant/FreeStyle\" rel=\"nofollow\">https://github.com/Blue2Giant/FreeStyle</a> 📦 Dataset: <a href=\"https://huggingface.co/datasets/Blue2Giant/FreeStyle_Dataset\">https://huggingface.co/datasets/Blue2Giant/FreeStyle_Dataset</a> ⚖️ Model Weights: <a href=\"https://huggingface.co/Blue2Giant/FreeStyle_Checkpoint\">https://huggingface.co/Blue2Giant/FreeStyle_Checkpoint</a> 📊 Benchmark: <a href=\"https://huggingface.co/datasets/Blue2Giant/FreeStyle_Bench\">https://huggingface.co/datasets/Blue2Giant/FreeStyle_Bench</a> 🔍 LoRA Metadata: <a href=\"https://huggingface.co/datasets/Blue2Giant/free_style_lora_meta\">https://huggingface.co/datasets/Blue2Giant/free_style_lora_meta</a>\n","updatedAt":"2026-06-19T02:17:21.666Z","author":{"_id":"64b914c8ace99c0723ad83a9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64b914c8ace99c0723ad83a9/B4gxNByeVY_xaOcjwiN1j.jpeg","fullname":"Wei Cheng","name":"wchengad","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":9,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.6955878138542175},"editors":["wchengad"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/64b914c8ace99c0723ad83a9/B4gxNByeVY_xaOcjwiN1j.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.20506","authors":[{"_id":"6a34a36c4c5c5e0d69bf1c03","name":"Jinghong Lan","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c04","name":"Wei Cheng","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c05","name":"Yunuo Chen","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c06","name":"Ziqi Ye","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c07","name":"Peng Xing","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c08","name":"Yixiao Fang","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c09","name":"Rui Wang","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c0a","name":"Yufeng Yang","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c0b","name":"Xuanyang Zhang","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c0c","name":"Xianfang Zeng","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c0d","name":"Difan Zou","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c0e","name":"Gang Yu","hidden":false},{"_id":"6a34a36c4c5c5e0d69bf1c0f","name":"Chi Zhang","hidden":false}],"publishedAt":"2026-06-18T00:00:00.000Z","submittedOnDailyAt":"2026-06-19T00:00:00.000Z","title":"FreeStyle: Free Control of Style-Content Dual-Reference Generation from Community LoRA Mining","submittedOnDailyBy":{"_id":"64b914c8ace99c0723ad83a9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64b914c8ace99c0723ad83a9/B4gxNByeVY_xaOcjwiN1j.jpeg","isPro":false,"fullname":"Wei Cheng","user":"wchengad","type":"user","name":"wchengad"},"summary":"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.","upvotes":14,"discussionId":"6a34a36c4c5c5e0d69bf1c10","projectPage":"https://blue2giant.github.io/FreeStyle/","githubRepo":"https://github.com/Blue2Giant/FreeStyle","githubRepoAddedBy":"user","ai_summary":"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.","ai_keywords":["LoRA mining","dual-reference generation","style transfer","content leakage","disentanglement mechanisms","attention-level enrichment constraint","frequency-aware RoPE modulation","Content Alignment Score","Rejection Score"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":2,"organization":{"_id":"643cb0625fcffe09fb6ca688","name":"Fudan-University","fullname":"Fudan University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6437eca0819f3ab20d162e14/kWv0cGlAhAG3iNWVxowkJ.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"64b914c8ace99c0723ad83a9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64b914c8ace99c0723ad83a9/B4gxNByeVY_xaOcjwiN1j.jpeg","isPro":false,"fullname":"Wei Cheng","user":"wchengad","type":"user"},{"_id":"68b04979f64bd1f33194cbcb","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/dtsZQPHpBFnOCmgHfSeNh.png","isPro":false,"fullname":"Chan","user":"yuUnuo","type":"user"},{"_id":"62f361e6231737ed2d741740","avatarUrl":"/avatars/f4a1053f9d9b3e703d138bc9753742c1.svg","isPro":false,"fullname":"huyaoqi","user":"yaoqi","type":"user"},{"_id":"62a6260e5ade0c7a3809ba14","avatarUrl":"/avatars/008e8d0e4dd0128ca7a326589ed34c73.svg","isPro":false,"fullname":"journey","user":"journeyStar","type":"user"},{"_id":"679f20275b562859c25a1bef","avatarUrl":"/avatars/240583b96f6fae5447612b329e93b99d.svg","isPro":false,"fullname":"sjbixiitu","user":"sjbixiitu","type":"user"},{"_id":"634bde123d11eaedd889e277","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1665916392312-noauth.png","isPro":false,"fullname":"Hengyuan Xu","user":"DobyXu","type":"user"},{"_id":"6682497fe365c0f666ff1149","avatarUrl":"/avatars/ba32c978761ef7ac8cc467184b8441a4.svg","isPro":false,"fullname":"Xinyao Liao","user":"leoisufa","type":"user"},{"_id":"6742f612924e80c3c81352d9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/Wmp_PY2t-CuJEoeCyR3NM.png","isPro":false,"fullname":"Haoling Xie","user":"HAOlingX","type":"user"},{"_id":"67761e674467879a54b4624a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/hbniDcGynaZIUZTNSgc2G.jpeg","isPro":false,"fullname":"Xiaoyun Yuan","user":"XiaoyunYuan","type":"user"},{"_id":"67da6acc05101e8e1d2c20a2","avatarUrl":"/avatars/1cfa3a1f59687db58af4e1b4a8767bfd.svg","isPro":false,"fullname":"Yang","user":"Yiying12","type":"user"},{"_id":"6586e61fd2ea3f329401777b","avatarUrl":"/avatars/262ae4b2535b5013c80171a31f0fb919.svg","isPro":false,"fullname":"te","user":"itachi3242","type":"user"},{"_id":"6343de25e01a38440ef02d5e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6343de25e01a38440ef02d5e/eumgZKT6vfTzINC6cYUrL.jpeg","isPro":false,"fullname":"xz","user":"frankzeng","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"643cb0625fcffe09fb6ca688","name":"Fudan-University","fullname":"Fudan University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6437eca0819f3ab20d162e14/kWv0cGlAhAG3iNWVxowkJ.png"},"query":{}}">

Papers

arxiv:2606.20506

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

Published on Jun 18

· Submitted by

Wei Cheng on Jun 19

Fudan University

Upvote

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.

Generated by Qwen/Qwen2.5-Coder-32B-Instruct

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.