Shilin-LU/TF-ICON

Leverages an off-the-shelf Stable Diffusion model without training or finetuning, using an "exceptional prompt" (information-free text) to improve real image inversion into latent space for accurate composition. The framework performs cross-domain image-guided composition by seamlessly integrating foreground objects into background contexts via diffusion-based inversion and synthesis. Supports both cross-domain and same-domain composition scenarios across diverse visual styles (photorealism, cartoon, oil painting) with customizable input masks and prompts.

821 stars. No commits in the last 6 months.