Image-to-LoRA-V2 - An open-source, training-free style transfer tool from the Moda community.

What is Image-to-LoRA-V2?

Image-to-LoRA-V2 (i2L-V2) is an open-source, training-free style transfer tool from the Moda community. It allows users to upload 1-8 style images and predict LoRA weights for a raw image model in a single inference. The tool supports three major frameworks: Z-Image, FLUX.2, and Hidream-O1, offering style fidelity superior to methods like IP-Adapter. LoRA images can be exported and reused, and it supports ControlNet integration for multi-style fusion and structural control.

Main functions of Image-to-LoRA-V2

• Training-free style LoRA generationUpload 1-8 images with a consistent style, and output a downloadable LoRA weight file directly from a single forward inference.
• Multi-base model adaptationIt natively supports three mainstream text-based image base models: Z-Image, FLUX.2, and Hidream-O1.
• Modular combination capabilityThe generated explicit LoRA can be used in conjunction with modules such as ControlNet, AttriCtrl, and Inpainting through standard interfaces.
• Multi-style fusionPredict a single LoRA from multiple reference images of different styles, allowing the generated image to inherit multiple visual attributes simultaneously.
• One-click upload and exportThe generated LoRA file can be uploaded to the Moda community with one click, or downloaded and used in the AIGC section and other tools.

The technical principle of Image-to-LoRA-V2

• Structured LoRA QueryModern diffusion Transformers have a large number of fully connected layers to be adapted, while the reference signals consist of only a few images, resulting in a significant scale mismatch. i2L does not generate all weights from a single pooling vector; instead, each LoRA query corresponds to a row or column of the LoRA matrix: for each adaptation layer, k queries generate rows of matrix A, and k queries generate columns of matrix B, for a total of 2kL queries. The reference image is encoded into an image token by a SigLIP2 encoder, concatenated with the LoRA query, and then fed into a single-stream Transformer for aggregation. Finally, each layer's independent compressed linear decoder reconstructs the LoRA matrix.
• Style-content decoupling:Ordinary image-text pairs can induce the model to encode reference semantics into LoRA as well (e.g., a reference image of a cat will cause the generated dog to also have the features of a cat). i2L is trained on MegaStyle-1M, constructing training tuples with consistent style but different content, and the prompt only describes the target content rather than the reference image, so that the loss reward is consistent with style and inhibits the copying of objects or identities as shortcuts.
• Asymmetric LoRA guidance:The reference style LoRA is applied to the positive branch of the CFG, while the neutral LoRA predicted from the pure gray map by the same i2L network is applied to the negative branch. The two branches are parameterized similarly, and their differences mainly reflect the style update brought by the reference map, allowing the guiding direction to amplify the style-related effects.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use Image-to-LoRA-V2

• Upload reference imageUpload 1-8 images with the target style to the Moda Creative Space. Mixing different themes helps improve generalization ability.
• Generate LoRA modelClick the "Generate" button, and the model will predict the LoRA weights through one forward inference and output them. .safetensors document.
• Input prompt wordsEnter the target content prompts and set the generation parameters such as image height, width, and LoRA intensity.
• Generate imageBased on the selected pedestal model and predicted LoRA, stylized images are generated in real time and the effects are previewed.
• Export and reuseUpload the LoRA file to the Moda community with one click, or download it and use it in tools such as AIGC Zone and ComfyUI.

Image-to-LoRA-V2's core advantages

• Ready to use without training: No need to run hundreds or thousands of training steps on a GPU; a single inference can generate a ready-to-use LoRA dataset.
• Leading in style fidelityIt surpasses baselines such as IP-Adapter and InstantStyle across multiple metrics including CLIP-Style, Aesthetic, PickScore, and HPSv2/v3.
• Strong content consistencyCLIP-Text scored the highest at 34.71, indicating that LoRA predictions effectively retained controllability of prompt content while applying a style.
• Modular and reusableOutput explicit LoRA rather than temporary conditional characteristics, which can be stored, interpolated, and reused, and can be combined with other control modules through the standard LoRA interface.
• Improved open source ecosystemThe model weights and code are completely open source, and users can experience them directly online in the Creative Space. The generated results can be seamlessly integrated into the Moda AIGC Zone.

Image-to-LoRA-V2 project address

• Model weights:https://modelscope.cn/collections/DiffSynth-Studio/Image-to-LoRA-V2
• Creative Space:
  • 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

Image-to-LoRA-V2 Comparison with Similar Products

Application scenarios of Image-to-LoRA-V2

• Brand visual unityQuickly extract brand design reference images to generate a series of marketing materials and social media images, ensuring consistent visual tone across all channels.
• Illustration style migrationConvert the style of an artist's reference work into reusable LoRA, and generate commercial illustrations of the same style in batches, avoiding hand-drawing one by one or repeated training.
• Game asset generation: Build a style library for game projects, and use LoRA combined with ControlNet to precisely control character and scene structure, accelerating concept design and asset iteration.
• E-commerce design efficiency improvementGenerate consistent detail page backgrounds, posters, and decorative elements based on product reference images, reducing design costs and improving listing efficiency.