The ControlNet paper defines what each condition map does

The image

The document at hand is not a single photographed frame but a technical paper and its companion code repository: "Adding Conditional Control to Text-to-Image Diffusion Models," posted to arXiv on 10 February 2023 by Lvmin Zhang, Anyi Rao and Maneesh Agrawala, and revised through November 2023. Alongside the prose, both the paper and the repository publish comparison panels: a source photograph reduced to a Canny edge map, a depth map, an OpenPose skeleton or a hand scribble, next to diffusion output generated from that map. These panels illustrate a method; they are not production stills from a shoot.

What the documents show

The paper states that ControlNet duplicates a pretrained diffusion model's weights into a locked copy that stays fixed and a trainable copy that learns the new condition, joined by "zero convolution" layers initialized at zero so training begins from the unmodified model rather than injecting noise. The abstract names edge maps, depth, segmentation and human pose as demonstrated conditions and states training held up on datasets from under 50,000 to over one million images. The repository adds detail the paper's abstract does not: it names nine conditioning types, including HED soft boundaries and normal maps derived from depth, and it discloses that an anime line-art model was withheld, in the maintainer's words, "pending risk evaluation" — a limit the source states about itself, not a claim added here.

Production context

This is an editorial reading beyond what either document states as production use: a depth or edge map pulled from a location photograph, a storyboard panel or a locked-off previsualization plate could hold a generated concept image to a specific camera setup before a shoot, while a text prompt supplies content the map cannot. The repository's own "Guess Mode," which it says lets "the ControlNet encoder try best to recognize the content of the input control map... even if you remove all prompts," suggests the structural map alone can carry a generation, a capability documented past what the original paper's abstract describes.

Reference versus imitation

Both documents describe conditioning on a computed structural map, not on a reference image's pixels or on a named artist's style. A workflow that credits ControlNet correctly says which map was extracted and from what source frame; one that claims the tool "recreates" a photograph exactly overstates what zero-convolution conditioning is documented to do. The repository's own decision to hold back a conditioning type over misuse risk is itself evidence that structural conditioning is understood, by its authors, as something that can be pointed at reproduction rather than mere reference.

• Which structural map was actually computed from the source, and does the pipeline record it?
• Is a claimed conditioning type named in the tool's own documentation, or assumed by whoever ran it?
• Would the same result be achievable from the text prompt alone, without the reference map?

A conditioning-type list is a first, checkable fact before any claim about how a look was held in place; the comparison panels in a paper or README remain examples of a method, not certified evidence from a production.