Image-to-Image

Image-to-image (img2img) is a generative technique where a diffusion model transforms an existing input image into a new output image, guided by a text prompt, a style reference, or both — rather than generating from pure noise as in standard text-to-image. Technically, instead of starting the denoising process from random noise, img2img starts from the input image with a controlled amount of noise added (the "denoising strength" or "strength" parameter, typically 0.3-0.8); a low strength preserves most of the original structure and just restyles it, while a high strength allows the model more freedom to deviate, approaching a fresh generation. Related conditioning techniques like ControlNet let builders lock specific structural elements (pose, depth map, edge outlines) while still letting the model reimagine color, texture, and style, giving product teams a dial between "light restyle" and "creative reinterpretation" rather than a single all-or-nothing setting. Why it matters for SaaS builders: img2img powers "restyle my photo" consumer apps, product-mockup generators (place a logo design onto a photographed t-shirt), architectural/interior-design visualization tools, and rapid concept-art iteration for game and film studios. It's cheaper and more controllable than pure text-to-image when the goal is "keep this, but change that," making it the workhorse behind most commercial AI photo-editing features. A concrete worked example — a real-estate SaaS offering "virtual staging" (furnishing empty rooms in listing photos): (1) agent uploads a photo of an empty living room and picks a staging style ("modern minimalist," "traditional," "Scandinavian"); (2) the app sends it to the img2img API with a prompt like "Add modern minimalist furniture: a grey sofa, wooden coffee table, and floor lamp, keep the room's walls, windows, and flooring unchanged" and `strength=0.55` — low enough to preserve the room's actual architecture; (3) the model uses a ControlNet depth map extracted from the original photo to preserve room geometry, wall angles, and window placement while adding furniture consistent with the actual perspective and lighting direction visible in the source photo; (4) the staged photo is returned in roughly 5-10 seconds and shown as a before/after toggle against the original so the agent can approve or regenerate with a different style; (5) approved staged images are watermarked "Virtually Staged" per most MLS and real-estate marketing regulations, which increasingly require explicit disclosure that furniture in a listing photo is not physically present. Builders serving regulated real-estate markets should treat this disclosure requirement as a hard product requirement, not an optional nicety.