Tight Inversion

Overview

Tight Inversion is a fresh way to edit images using text-to-image diffusion models. Instead of relying only on text prompts it conditions directly on the image itself. This means better reconstructions and more flexible edits.

A team from Tel Aviv University and Snap Research put this together. The group includes Edo Kadosh Nir Goren Or Patashnik Daniel Garibi and Daniel Cohen-Or. They’ve worked on other big advancements in AI-based image editing.

Editing real images with AI has always been tricky. If you focus on keeping details sharp it’s hard to make big edits. If you push for flexibility you lose fine textures. Tight Inversion fixes this by conditioning on the image itself instead of relying on text alone.

• Better Reconstruction. Keeps details like reflections tattoos and textures.
• More Control. Works with tools like Prompt2Prompt and LEDITS++ for precise edits.
• Works with Multiple Models. Supports Stable Diffusion XL SDXL-Turbo and Flux.
• Adjustable Strength. Users can tweak how much the original image affects edits.
• Plug-and-Play. Can work with methods like DDIM Inversion ReNoise and RF-Inversion.

How Does Tight Inversion Work?

• Image-Based Conditioning. Instead of relying on text it conditions directly on the input image.
• Image Adapters. Uses tools like IP-Adapter and PuLID to help the model understand images better.
• Better Noise Control. Tweaks the inversion process so outputs match the original more closely.