A small adapter claimed to steer a much larger frozen model

The image

The record is the paper "T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models," posted to arXiv on 16 February 2023 by Chong Mou and seven co-authors, revised on 20 March 2023, and its companion repository. Both publish side-by-side panels of a sketch, a depth map or a keypose skeleton next to a full-color generation conditioned on it — illustrations of a method rather than documentation of any actual advertising or film shoot.

What the documents show

The paper's abstract states the goal directly: to "learn simple and lightweight T2I-Adapters to align internal knowledge in T2I models with external control signals, while freezing the original large T2I models," aiming at "rich control and editing effects in the color and structure of the generation results." The repository names the specific adapters released — sketch, canny edge, line art, OpenPose keypoints, depth (via two named estimators) and color — and reports a scale comparison the abstract does not spell out: a T2I-Adapter for Stable Diffusion XL uses 77 million parameters to steer a 2.6-billion-parameter base model, which the repository calls the "Original Recipe" carried over from the smaller SD-1.4/1.5 adapters. Neither document names ControlNet directly in the passages read here; the comparison offered is against full model fine-tuning, not a named competing adapter method.

Production context

Read editorially, an adapter documented at roughly 3 percent of the base model's parameter count matters operationally where storage or retraining cost is a constraint — a studio holding several house styles or client palettes as separate small adapter files rather than several full model copies. The paper's stated freezing of the base model is the documented reason multiple adapters can be swapped against one shared backbone, per its own description, though neither source documents a specific production that did this.

Reference versus imitation

The adapters condition generation on computed signals — edges, a depth estimate, keypoints, or a color layout — not on an artist's named style or a specific copyrighted image's pixels. Describing T2I-Adapter as "the same as ControlNet" goes beyond what either document states here, since the two projects are documented separately and this entry's sources make no direct functional comparison between them; describing it only as "a lightweight external control signal," in the paper's own words, stays inside what is documented.

• Which named adapter (sketch, depth, keypose, color) is being used, and does the source specify the estimator behind it?
• What fraction of the base model's parameters does the adapter add, per the repository's own figures?
• Does any cited source directly compare this method to ControlNet, or is that comparison being assumed?

A parameter-count comparison from the authors' own repository is a specific, checkable claim; treating two separately developed adapter methods as interchangeable without a source that compares them directly is not.