shilin-lu

Professional software vendor delivering innovative solutions on the Softono platform. Specialized in both open-source and proprietary software development.

Visit Website

Total Products

Software by shilin-lu

Open Source

TF-ICON

TF-ICON is an official implementation of a novel Training-Free Image COmpositioN framework presented at ICCV 2023. It enables cross-domain image-guided composition by seamlessly integrating user-provided objects into specific visual contexts using text-driven diffusion models. Unlike existing diffusion-based methods that require costly instance-based optimization or model finetuning, TF-ICON leverages off-the-shelf models without any additional training or parameter updates. A core innovation is the exceptional prompt, an empty prompt designed to enhance the accurate inversion of real images into latent representations, which forms the basis for high-quality compositing. The system has been tested with Stable Diffusion and demonstrates superior performance iniverse visual domains compared to state-of-the-art inversion methods and prior baselines on datasets including CelebA-HQ, COCO, and ImageNet. This tool is ideal for researchers and developers seeking advanced image editing capabilities without the computa

AI & Machine Learning ML Frameworks

819 Github Stars

Open Source

MACE

MACE is an official implementation of the CVPR 2024 paper Mass Concept Erasure in Diffusion Models. It is a fine-tuning framework designed to prevent text-to-image diffusion models from generating harmful, misleading, or unwanted content by removing specific concepts upon request. While existing methods struggle to handle more than a few concepts simultaneously without degrading the model, MACE successfully scales to erase up to 100 concepts at once. The system achieves an effective balance between generality, by erasing concept synonyms, and specificity, by preserving unrelated content. This is accomplished through closed-form cross-attention refinement and specialized LoRA fine-tuning that eliminates undesirable information without interference. MACE also employs a unique integration technique that combines multiple LoRA modules without causing catastrophic forgetting. Extensive evaluations demonstrate that MACE outperforms prior methods across four key tasks: object erasure, celebrity erasure, explicit con

AI & Machine Learning Mobile Development ML Frameworks

5K Github Stars