Cutoff is an open-source extension for the Stable Diffusion WebUI that limits how far individual prompt tokens influence an image, fixing the classic problem of attributes bleeding between subjects. The extension has about 1,200 stars on GitHub and supports SD 1.x, SD 2.x and SDXL models.
Anyone who has prompted “a woman in a red dress standing next to a blue car” and received a blue dress or a red car has met token bleeding: in Stable Diffusion, every word tends to affect the whole picture. Cutoff attacks this by restricting the influence scope of chosen target tokens. You list the words to contain, and the extension pads their embeddings outside the relevant context and interpolates between padded and original vectors, with adjustable weight and a choice of Lerp or SLerp interpolation. The result is cleaner attribute binding in multi-subject scenes without changing models or rewriting prompts into fragments.
Because Cutoff works at the conditioning level inside the generation process, it composes with the rest of the WebUI ecosystem, including LoRAs and ControlNet, and its settings can be tuned per generation without restarting anything. It is especially valued in multi-character and anime generations, where two characters with different hair and outfit colors are a classic failure case for vanilla prompting.
Cutoff is free and open source, distributed through GitHub. There is no paid version; you install it into your own Stable Diffusion WebUI setup at no cost.
Cutoff is for Stable Diffusion users who generate scenes with several subjects or precise color and outfit requirements and are tired of attributes leaking across the image.