VQGAN-CLIP (often written VQGAN+CLIP) is an open-source method for generating and editing images from a text prompt, developed by members of EleutherAI. It is described in a 2022 paper by Katherine Crowson, Stella Biderman and five co-authors, accepted at the ECCV 2022 conference. The method joins two existing neural networks: VQGAN, which produces the image, and CLIP, a multimodal encoder (a model that reads both images and text) that guides VQGAN toward the prompt.
Because CLIP does the guiding, the method needs no training for the task. The paper reports images of high visual quality from complex prompts, and higher visual quality than earlier approaches such as DALL-E, GLIDE and Open-Edit.
The code is in a public GitHub repository under the MIT license. A user installs the Python dependencies, downloads a VQGAN checkpoint (the saved weights of a trained model, for example one trained on ImageNet) and runs a script with a prompt such as "A painting of an apple in a fruit bowl".