Reviewing image enhancement in ComfyUI without confusing detail with accuracy
An enhancement stage can add convincing texture while changing the scene. Decide whether the task is simple resizing, perceptual enlargement or correction of a particular defect before choosing a workflow, and preserve the unmodified source for comparison.
Choose the operation that matches the defect
If the image only needs a different pixel size, an ordinary resize may be sufficient. If it needs perceptual enlargement, inspect a maintained model-based upscale example. If an object or region is wrong, consider a targeted edit rather than expecting enlargement to correct semantic mistakes.
Write the acceptance test in visible terms: legible lettering, an unchanged product outline, consistent skin texture or a clean edge. Avoid a vague requirement for more detail. A model can invent plausible texture that makes the image less faithful to the original.
Use a documented graph
The official ComfyUI upscale example distinguishes the image input from the loaded upscale model. Follow its actual connections instead of connecting an image to a model-loader input or relying on an invented node setting. Preserve the original workflow before adding the enhancement branch.
For a targeted correction, use the inpainting documentation and inspect the mask carefully. Save the mask and input image with the output. The edit should be evaluated in context, including its boundary with the untouched area, rather than only through a cropped preview.
Compare at delivery scale
Review the original and processed images at the size at which the audience will see them, then inspect suspicious regions more closely. Check repeating patterns, typography, faces, product markings and fine lines. A sharpened preview on one screen is not a complete delivery check.
Keep native generation dimensions and final dimensions as separate metadata. If a later generative pass changes the content, disclose that process in the project record. Do not describe invented details as recovered information, especially where the image will be used as factual evidence.
Measure the entire enhancement route
Record model loading, processing and saving alongside peak memory if performance matters. Compare accepted outputs, not just the fastest preview. Test optional stages independently and include failures. An enlargement route that works once but regularly exhausts memory may not be suitable for a repeatable production task.
Retain the source, graph, masks, model identities and final file together. This revision replaces unsourced quality comparisons and incorrect wiring suggestions with a documented review process. It does not declare one upscale model universally superior or claim a new measured gain in image quality.