A ComfyUI image-to-image workflow is useful when you like the composition of an image but need a controlled variation. The challenge is deciding which parts may change and which parts must remain recognizable.
Write an edit brief first
For a fictional creator portrait, a clear brief might be: “Keep the face, camera angle and jacket shape; change the scene from a plain wall to a softly lit studio.” That gives you concrete acceptance criteria. “Make it better” gives neither you nor the generator a useful boundary.
If the requested change affects only a small region, consider inpainting instead. Image-to-image can change the whole picture, including areas you did not intend to revise.
Load a model-specific example
The official ComfyUI tutorial demonstrates a reference image feeding an image-to-image graph. Load its example, select the required checkpoint and provide the input through Load Image. Follow the exact model-family example rather than assuming an SD1.5 graph transfers unchanged to every newer architecture.
Understand denoise without treating it as a magic identity lock
In the basic KSampler workflow, lower denoise generally changes less of the input, while higher denoise permits larger changes. At full denoise, the original latent information is effectively replaced by noise. This behavior is explained in the official tutorial; reference-conditioned systems can add other controls.
Use that knowledge to design a small comparison, not to pick a universal “best” number. Keep the model, seed, prompt and dimensions fixed, then test several denoise levels supported by the graph. Label the outputs before comparing them so a preferred image does not become an undocumented accident.
A three-candidate edit test
| Candidate | Question |
|---|---|
| Conservative change | Did the requested edit happen at all? |
| Moderate change | Is the edit visible while the important structure survives? |
| Stronger change | Did the result cross the boundary into a different character or composition? |
These are comparison labels, not fixed parameter presets. Reject a technically attractive output if it no longer matches the creator or product. The acceptance brief should win over the urge to keep the most dramatic image.
Example: move a creator into a new setting
Begin with a neutral chest-up portrait of an original adult character. Keep the face unobstructed and the lighting easy to read. For the scene variation, ask for an indoor reading corner with warm daylight and a simple background. Do not also request a new haircut, a profile angle and a different facial expression on the first attempt.
Compare eye spacing, jaw outline, hairline and the relationship between the head and shoulders. Check the jacket seams and any small accessories. If the face changes while the background improves, record that tradeoff and try a more localized method or a reference-based workflow.
When img2img is not enough for consistent characters
Image-to-image can retain useful structure, but keeping a recurring identity across large pose and location changes is a separate production problem. Use an approved reference set and evaluate an identity-specific method appropriate to the model. The ComfyUI character guide explains how to test that without confusing scene resemblance with identity consistency.
Save the result as an edit history
Keep the original, each accepted intermediate and the graph version. If you repeatedly feed an edited image into another edit, small changes can accumulate. Compare later versions against the original reference, not only the immediately preceding image.
Finish with the correct crop and, if needed, a separate upscale pass. If you prefer a persona-led hosted workflow for new images and videos, create your character in Clout and use the same written acceptance brief for the resulting batch.
