Img2img and Reference Images in Practice: Controlling Composition and Character Consistency

Img2img means uploading an image as the base canvas: the model repaints on top of it and outputs a new picture with a structure close to the original. Using a reference image goes a step further: the model treats the reference as visual evidence for locking in a character's face, an art style or a composition, rather than just swapping the background. The two are often combined, and together they form the two basic moves of working with reference images.

As a rule of thumb, use img2img when you want to keep the skeleton of an image and change only part of it; rely on a reference image when you want to take the person in that picture and generate new scenes in batches. Pure text-to-image has no visual anchor and depends entirely on luck with words; once a reference image is involved, results shift from a random draw to controlled adjustment — the foundation for every consistency technique below.

Img2img usually exposes a denoising strength parameter — some platforms call it similarity or redraw strength — that decides how far the new image may drift from the original. Lower values keep the output close to the source; higher values give the model more freedom, at the cost of the person and fine details falling apart more easily.

The general logic: use medium strength for light retouching (new background, different lighting, a change of clothing color); use low strength when pose and composition must survive (the gesture and camera angle cannot move); use high strength only for drastic rework that borders on repainting. There is no universal number — value ranges and defaults differ across models and platforms, so start from the middle, nudge in small steps, and watch which way the picture moves before adjusting further. Two classic failures when you overshoot: strength too low and the background never changes; too high and your character comes back with a different face.

The most common request is same person, new background — moving an indoor photo to the seaside, or turning day into dusk. Such tasks generally call for medium strength: the model gets enough room to repaint the scenery without touching facial features or body shape.

A more reliable approach adds inpainting: mask out the person to protect them, apply bigger changes only to the background region, and keep the person at low strength or exclude them from repainting entirely. The background can then be replaced boldly while the person barely moves. Most mainstream img2img tools usually support masks in one form or another, and the workflow is similar everywhere: paint over the area you want changed, describe the new background, and generate.

When pose and composition are the whole point — producing images from a fixed template, or keeping a consistent camera angle across a set — use low strength. At low strength the model mostly resurfaces the picture: lighting, texture and color shift, while gestures, limb positions and layout stay put. This kind of low-strength pass is also a common way to unify the color mood of an image set.

Low strength has a cost: it struggles to change content. If you want to swap clothes or hair while keeping the pose, low strength often will not move; either raise it a little, or use inpainting on just the part you want changed. The test is simple: check whether the output looks like a close relative of the original. Too similar? Add a bit of strength. The person starts to distort? Subtract some.

Putting the same person into multiple images is the most valuable — and most failure-prone — use of reference images. The common recipe has three parts: fix the random seed, fix the character description, and reuse the same reference image.

A fixed seed gives the model the same random starting point; the description of the character's appearance (hairstyle, face shape, clothing details) must stay word-for-word identical every time; and the same character reference is layered on top. The three constraints reinforce each other, and cross-image similarity is usually far better than text alone. Many tools generally also support reference weighting or multiple reference images, so a face reference and a style reference can be supplied separately, with slightly more weight on the face. It also helps to lock only one character at a time; when several characters must appear together, keep each one's reference matched to its own figure, otherwise their faces tend to bleed into each other. Even then, small differences between images remain normal for current technology: generate several at once, keep the closest one, and if necessary iterate with it as the new reference.

Image-to-video tools generally support a first-frame approach: you upload one image as the opening frame, and the picture comes alive from there. The quality of that frame decides how the whole clip feels, so the usual workflow is to polish the first frame with img2img first — composition, character and lighting all approved — before handing it to the video model for motion. The first frame can, of course, also be produced by text-to-image directly and then fed into the video step.

When writing the prompt for a first-frame job, focus on how things move: camera push-ins, character actions, changes in the environment — not on re-describing what is already in the picture. Some tools usually also support a first-and-last-frame mode: you supply one more image as the ending frame and the model fills in the transition, which suits scene cuts or looping animation.

With a reference image in play, text has not lost its job — it has changed roles: the reference answers what things look like, the text answers what to change and what you want. Well-matched writing keeps the preserved subject brief, because the image already carries it, and spends its words on what changes: the new background, new lighting, new action, new style. For example, given a reference of a standing figure, a prompt like the same person, seaside at dusk, backlit, coat caught by the wind is usually far more effective than a long paragraph re-describing the face.

Two mistakes are common. First, text that contradicts the reference — short hair in the picture, long hair in the prompt — leaves the model torn and the output messy. Second, text that describes the scene in exhaustive detail ends up fighting the reference's guidance. The rule of thumb: do not repeat what the reference already shows; add only information the image lacks, and say clearly what should change and what must stay.