AI image generation includes several related operations, but their names describe different starting conditions. A system can begin with a description, an existing picture, a selected region, or an expanded canvas. Understanding those distinctions explains what information enters the process and what kind of result comes back. This reference covers four method names and the documented mechanics behind them. Hugging Face Diffusers supplies concrete examples of diffusion pipelines; Adobe Photoshop supplies an example of canvas expansion. Their controls are implementation-specific, rather than universal rules for every image model. The guide does not compare generated imagery with photography or recommend a production approach. It separates the vocabulary of generation from decisions about commissioning creative work. The linked Diffusers development documentation was checked on 17 October 2026; controls below describe that dated documentation, not a stable cross-version specification.
What the method names describe
Text-to-image and image-to-image distinguish starting inputs. The first uses a text description; the second also receives an initial image. Inpainting and outpainting distinguish the region being generated: inside a selected area or beyond an existing image boundary. A single editing workflow can therefore involve more than one method. Diffusers documents text-conditioned generation, image-conditioned generation, and masked editing separately. Hugging Face: Text-to-image Hugging Face: Image-to-image Hugging Face: Inpainting
These labels do not identify a particular model architecture or file format. Here, diffusion is the documented implementation used to explain the operations, not a claim that every product follows identical internals. A hypothetical sequence might generate a landscape from text, transform its appearance using that image as input, replace a selected object, and extend the surrounding canvas. The labels describe those operations, not four independent guarantees about the resulting scene.
Text-to-image: description to generated image
In the Diffusers explanation, text-to-image starts with a prompt and random initial noise. The model progressively removes noise under the guidance of the prompt. After the configured denoising steps, the representation is decoded into an image. The prompt supplies descriptive conditioning; it is not an existing photograph that the system simply retrieves and returns. Hugging Face: Text-to-image
The documented pipeline loads a model checkpoint and receives a prompt. Its output includes generated images. Model architecture and training influence results, so identical wording is not a specification that different models must interpret identically. A hypothetical prompt describing a red sphere on a grey pedestal identifies requested subject matter and appearance; it does not prescribe the exact value of every output pixel. Additional conditioning mechanisms are separate capabilities rather than something implied by the phrase text-to-image. Hugging Face: Text-to-image
Image-to-image: transforming an initial image
Image-to-image adds an existing picture to the prompt. In the latent-diffusion process described by Diffusers, the initial image is encoded into a latent representation, noise is added, and the model denoises that representation with prompt conditioning. A decoder then produces the output image. The input contributes visual structure, but the process is generation rather than a promise to preserve every original feature. Hugging Face: Image-to-image
Diffusers exposes a strength parameter controlling how much noise is added to the initial image. Greater strength permits more transformation; at its documented maximum, the process effectively ignores the starting image. This meaning belongs to the documented pipeline: another product may use different scales or controls. Image-to-image alone does not identify a local editing area, because an initial image and an edit mask are different inputs. Hugging Face: Image-to-image
Inpainting: generating within a mask
Inpainting replaces or edits selected areas of an image. The Diffusers example takes an image, a mask, and a prompt. In that implementation, white mask pixels identify areas to generate and black pixels identify areas to retain. The convention matters because reversing the mask reverses the intended editing region. Mask colours and selection interfaces must be interpreted according to the particular tool. Hugging Face: Inpainting
A hypothetical mask might cover an empty tabletop where a new object is requested. The image supplies surrounding context, the mask supplies location, and the prompt describes the requested content. Diffusers also documents that some pipelines can change unmasked regions; restoring original unmasked pixels is a separate compositing operation. A selection therefore expresses the intended region, but exact preservation outside it is not an automatic property of every inpainting implementation. Hugging Face: Inpainting
Outpainting: extending the canvas
Outpainting generates content beyond an existing image boundary. Adobe demonstrates this operation through Photoshop Generative Expand: the user drags Crop tool handles outward to enlarge the canvas, selects Generative Expand, and generates a variation. Adobe also documents a blank-prompt workflow in which Photoshop extends the image without an additional written description. The feature name is Adobe-specific; the general operation is boundary expansion. Adobe: Generative Expand
For a hypothetical square landscape extended into a wider rectangle, the new side regions contain generated material rather than previously hidden photographic evidence. This distinguishes expansion from cropping, which removes part of the visible area, and from ordinary resizing, which changes dimensions. Inpainting and outpainting can share masked-generation techniques, but their spatial purpose differs: one fills a region within the working picture, while the other creates material beyond its earlier edges.
Inputs, controls, and output properties
A prompt, initial image, and mask carry different information. The prompt describes content; the initial image contributes visual information; the mask identifies an editing region. Diffusers additionally documents guidance scale, inference steps, and strength. These settings affect the generation process but are not synonyms for the method names. A product can expose some controls and hide others behind its interface. Hugging Face: Text-to-image Hugging Face: Image-to-image Hugging Face: Inpainting
Guidance scale controls the influence of prompt conditioning in the documented pipelines, while inference steps specify denoising iterations. Neither is a pixel dimension. Likewise, an instruction containing a resolution phrase is not the same as a configured output-size parameter. The dimensions, transparency, and encoding of an exported image are separate characteristics, defined in Creative Asset Specifications, rather than inherent meanings of text-to-image or inpainting. Hugging Face: Text-to-image
Randomness and documented limitations
Diffusion generation is stochastic: randomness is part of how different outputs are produced. Diffusers documents generators and seeds for controlling random values, along with deterministic algorithms where available. It also warns that identical seeds do not guarantee reproducibility in every environment. CPU and GPU random-number generation and computational settings can affect results. Hugging Face: Reproducibility
Consequently, a seed is one component of a generation setup, not a complete portable image recipe. A method label also leaves the checkpoint, conditioning, scheduler, and implementation unspecified. The documented distinction is between controlling randomness and guaranteeing identical pixels across environments. This reference defines the operations and their controls; it does not turn those controls into promises of accurate objects, identical characters, or unchanged compositions. Hugging Face: Reproducibility Hugging Face: Image-to-image
Frequently asked questions
Does text-to-image require an input photograph?
Not in the basic Diffusers text-to-image workflow. It begins with a prompt and random noise. Image-conditioned capabilities add other inputs and should not be confused with that basic definition. Hugging Face: Text-to-image
Is image-to-image the same as inpainting?
No. Image-to-image uses an initial image; inpainting additionally identifies an editing region with a mask in the documented workflow. The concepts can be combined within a larger pipeline. Hugging Face: Image-to-image Hugging Face: Inpainting
What does a white inpainting mask mean?
In the Diffusers documentation, white marks the area to fill and black marks the area to keep. This is an attributed convention, not a claim about every editing tool. Hugging Face: Inpainting
Can outpainting work without a written prompt?
Yes, in Adobe Photoshop’s documented Generative Expand workflow. Adobe describes leaving the prompt blank and generating an extension using the existing image as context. Other tools may expose different requirements. Adobe: Generative Expand
Will using the same seed always recreate an image?
No. Diffusers documents ways to control randomness but explicitly notes that reproducibility is not guaranteed even with an identical seed. The computation environment and other pipeline settings also matter. Hugging Face: Reproducibility