T2I-Adapter

T2I-Adapter, from Tencent ARC, adds structural control to text-to-image diffusion models through small adapter networks instead of retraining or duplicating the base model. Each adapter learns to translate one kind of guidance signal, such as a sketch, Canny edge map, line art, keypoint pose, depth map from MiDaS or ZoeDepth, or color layout, into internal conditioning that steers the diffusion model during generation. The adapters weigh in at roughly 77 to 79 million parameters, a small fraction of SDXL's 2.6 billion, which makes them cheaper to train and faster at inference than ControlNet while producing comparable control quality. Official adapters cover both Stable Diffusion 1.x and SDXL, with native support in Hugging Face Diffusers and public demos including Stability AI's Stable Doodle sketch tool built on the technique. Released under the Apache 2.0 license, it suits developers and researchers who need spatial or stylistic control over image generation without heavy computational overhead.