Image-to-LoRA-V2 - A Training-Free Style Transfer Tool Open-Sourced by ModelScope Community

What is Image-to-LoRA-V2

Image-to-LoRA-V2 (i2L-V2) is a training-free style transfer tool open-sourced by the ModelScope community. Upload 1-8 style images, and predict the LoRA weights of a text-to-image model in a single inference. The tool supports three major bases: Z-Image, FLUX.2, and Hidream-O1, with style fidelity surpassing methods such as IP-Adapter. The LoRA can be exported and reused, and supports combination with ControlNet to achieve multi-style fusion and structural control.

Main Features of Image-to-LoRA-V2

• Training-Free Style LoRA Generation: Upload 1-8 style-consistent images, and directly output a downloadable LoRA weight file in a single forward inference.
• Multi-Base Model Adaptation: Natively supports Z-Image、FLUX.2, Hidream-O1, three mainstream text-to-image base models.
• Modular Composition Capability: The generated explicit LoRA can be used in combination with modules such as ControlNet, AttriCtrl, and Inpainting through standard interfaces.
• Multi-Style Fusion: Predict a single LoRA from multiple reference images of different styles, allowing the generated image to simultaneously inherit multiple visual attributes.
• One-Click Upload and Export: The generated LoRA can be uploaded to the ModelScope community with one click, or downloaded and further used in the AIGC Zone and other tools.

Technical Principles of Image-to-LoRA-V2

• Structured LoRA Query: Modern diffusion Transformers have a large number of fully connected layers to be adapted, while the reference signal consists of only a few images, resulting in a clear scale mismatch. i2L does not generate all weights from a single pooled vector; instead, each LoRA query corresponds to a certain row or column of a LoRA matrix: for each adaptation layer, k queries generate the rows of matrix A, and k queries generate the columns of matrix B, with a total number of queries of 2kL. Reference images are encoded into image tokens by the SigLIP2 encoder, concatenated with LoRA queries and fed into a single-stream Transformer for aggregation, and then the LoRA matrix is reconstructed by a per-layer independent compressed linear decoding head.
• Style-Content Disentanglement:Ordinary image-text pairs will induce the model to also encode reference semantics into the LoRA (e.g., a reference image of a cat makes a generated dog also carry cat features). i2L is trained on MegaStyle-1M, constructing training tuples with consistent style and different content, and the prompt only describes the target content rather than the reference image, so that the loss rewards style consistency and suppresses copying objects or identities as shortcuts.
• Asymmetric LoRA Guidance:The reference style LoRA acts on the CFG positive branch, while a neutral LoRA predicted by the same i2L network from a pure gray image acts on the negative branch. The two branches have similar parameterization, and their difference mainly reflects the style update brought by the reference image, allowing the guidance direction to amplify style-related effects.

Follow WeChat and reply "open source" to join theAI Open Source Project Communication Group

How to Use Image-to-LoRA-V2

• Upload Reference Images: Upload 1-8 images with the target style in the ModelScope Studio; mixing different themes helps improve generalization ability.
• Generate LoRA Model: Click the generate button, and the model predicts LoRA weights through a single forward inference and outputs .safetensors files.
• Enter Prompt: Fill in the target content prompt, and set generation parameters such as image height, width, and LoRA strength.
• Generate Image: Generate stylized images and preview the results in real time based on the selected base model and predicted LoRA.
• Export and Reuse: Upload the LoRA file to ModelScope with one click, or download it and continue to use it in tools such as the AIGC Zone and ComfyUI.

Core Advantages of Image-to-LoRA-V2

• Training-Free and Ready to Use: No need to run hundreds or thousands of training steps on a GPU; a single inference generates a directly usable LoRA.
• Leading Style Fidelity: Comprehensively surpasses baselines such as IP-Adapter and InstantStyle across multiple metrics including CLIP-Style, Aesthetic, PickScore, and HPSv2/v3.
• Strong Content Consistency: CLIP-Text score reaches as high as 34.71, indicating that the predicted LoRA applies the style while better preserving prompt content controllability.
• Modular and Reusable: Outputs an explicit LoRA rather than temporary conditional features; it can be stored, interpolated, and reused, and can be combined with other control modules through the standard LoRA interface.
• Complete Open-Source Ecosystem: Model weights and code are fully open source, the Studio can be experienced directly online, and generated results can be seamlessly integrated into the ModelScope AIGC Zone.

Project Address of Image-to-LoRA-V2

• Model Weights：https://modelscope.cn/collections/DiffSynth-Studio/Image-to-LoRA-V2
• Studio：
  • Z-Image Version：https://modelscope.cn/models/DiffSynth-Studio/ZImage-i2L-v2 
  • FLUX.2 Version：https://modelscope.cn/models/DiffSynth-Studio/KleinBase4B-i2L-v2 
  • Hidream-O1 Version：https://modelscope.cn/models/DiffSynth-Studio/HidreamO1-i2L-v2 

Comparison of Image-to-LoRA-V2 with Similar Competing Products

Application Scenarios of Image-to-LoRA-V2

• Brand Visual Consistency: Quickly extract the style of brand design reference images, generate serialized marketing materials and social media images, and ensure consistent visual tone across all channels.
• Illustration Style Transfer: Convert the style of an artist's reference works into a reusable LoRA, batch-generate commercial illustrations in the same style, and avoid drawing each one by hand or repeatedly training.
• Game Asset Generation: Establish a style library for game projects, precisely control character and scene structure through LoRA combined with ControlNet, and accelerate concept design and asset iteration.
• E-commerce design efficiency improvement: Generate style-consistent detail page backgrounds, posters, and decorative elements based on product reference images, reducing design costs and improving listing efficiency.