What changes when you add an image
A text-only generation has to invent structure. Once you supply an image, structure is already decided, and the model's job narrows to reinterpreting it. This is why i2i is the workhorse of production work while text-only generation is the ideation tool: you can predict roughly what will come back.
Under the hood, most implementations add noise to your image and then denoise it toward the prompt. How much noise gets added is the strength setting, and it is the difference between a color grade and a completely new picture.
The strength dial is most of the interface
| Strength | What survives | Use for |
|---|---|---|
| Low (0.1 to 0.3) | Almost everything | Grades, small cleanups, texture passes |
| Medium (0.4 to 0.6) | Layout, pose, palette | Restyling, medium changes, look development |
| High (0.7 to 0.9) | A rough composition | Loose reinterpretation, concept variants |
Everyone's first mistake is running high strength and being surprised the source is gone. Everyone's second mistake is running low strength and concluding the model cannot restyle. Find the value where structure holds and style moves, note it, and reuse it for the whole shot.
Picking the right task
- Change the whole frame's look. That is i2i.
- Change one object or region. Inpainting, with a mask.
- Apply another picture's aesthetic. Style transfer, which is i2i with a style reference rather than a text instruction.
- Extend the frame outward. Outpainting.
- Fix the wrong thing at the wrong resolution. Do the repair first, then upscale.
Choosing the narrowest task that can do the job is the single most useful habit in this workflow. Every extra pixel you allow the model to touch is a pixel it can get wrong.
Multiple reference images
Newer models accept several images at once, each playing a role: one for composition, one for a character's identity, one for a product or garment. This is where image to image stopped being a style tool and became an assembly tool. Two rules make it work:
- Say what each image is for in the prompt. Without that, the model guesses, and it usually guesses that you wanted an average.
- Keep the references clean and cropped tight on the thing that matters. A busy reference contributes its clutter along with its subject.
Prompting an i2i pass
Describe the change, not the picture. The picture is already in front of the model, and re-describing it competes with the pixels you supplied.
- ✅
repaint as hand-inked comic art, flat shadows, keep pose and camera angle - ⚠️
a woman standing by a window in a blue dressrestates what the source already fixes
Two failure modes to expect. Text and logos degrade on every pass, because they are re-synthesized each time; mask them out if they matter. And quality drifts if you chain passes, since each run inherits the previous run's artifacts. When you find yourself on the fourth pass over the same file, go back to the original and do the work in one considered step instead.