FreeStyle mines community LoRAs to improve style-content dual-reference image generation
jiqizhixin · x · 2026-07-27
Researchers from Fudan University, StepFun, Westlake University, and the University of Hong Kong present FreeStyle, a method for dual-reference image generation that tries to match both a style reference and a content reference without leaking one into the other.
- They mine community-created LoRAs to build a large dataset of clean style/content pairs.
- The training pipeline uses a two-stage setup plus attention and frequency tricks to reduce style leakage while keeping content details intact.
- In style transfer and dual-reference benchmarks, FreeStyle reports a strong balance across style alignment, content preservation, and leakage suppression, outperforming prior methods.
- The paper is titled “FreeStyle: Free Control for Style–Content Dual-Reference Generation from Community LoRA Mining.”
More from Multimodal
- Maginary adds GPT-Images2 support and teases MCP and x402 payments — lutian · 2026-07-27
- RunwayML lets a creator finish a music video in 4 hours from a blank slate — notiansans · 2026-07-27
- Can a LoRA make image models generate ordinary-looking people instead of perfect faces? — baben7 · 2026-07-27
- ComfyUI user looks for a simpler workflow to edit photos and pose-driven images — jd142 · 2026-07-27
- Algorithmic fern artwork shows one frond across 26 ages, built with numpy and PIL — repligate · 2026-07-27
- LinkSpotlight adds Alt+H focus mode to untangle complex ComfyUI graphs — dingsl771 · 2026-07-27