FreeStyle: Dual-Reference Image Generation

机器之心 · wechat · 2026-07-17

This article introduces the image generation paper FreeStyle, focusing on the dual-reference generation task combining a "content reference image + style reference image".

The article points out that the difficulty lies in content leakage: when learning style, models easily copy over the subjects, scenes, and local details from the style image. To address this, the authors propose a pipeline to mine data from community LoRAs: filtering usable content and style LoRAs, then generating, verifying, and combining them to build two types of training data—around 619,000 SRef style reference entries and 480,000 CRef+SRef dual-reference entries, covering a massive range of style categories.

Methodologically, FreeStyle employs two-stage training: the first stage targets style-reference generation, introducing an attention-level enrichment constraint to suppress content leakage from the style image; the second stage targets content-style dual-reference generation, introducing frequency-aware RoPE modulation to reduce local patch-level correspondence between the style and output images, thereby lowering the risk of copying specific content. The article also details a constructed benchmark evaluating style consistency, content consistency, instruction adherence, aesthetic quality, and leakage degree, alongside releasing the project code, datasets, and model weights.

Original post →

More from Multimodal

Multimodal channel →