Give ComfyUI a picture, not your lighting data.

A render leaves V-Ray with recoverable window detail. It enters a node graph, acquires nicer planting and comes back with a white rectangle where the sky used to be. The foliage improved. The exposure range resigned.

This week's ComfyUI questions ask for workflows that enhance architectural renders while preserving the design. Most answers begin with models, samplers and denoise. Color arrives late, usually after something has turned gray, radioactive or oddly crunchy.

The fix is procedural. Draw a clear boundary between scene data and the image the model is meant to see.

Why should the linear EXR stay outside the AI pass?

A renderer's scene-linear file is built to hold light values, including values brighter than a display can show directly. OpenColorIO defines scene-referred color around the light captured or generated by a scene, while display-referred values describe an image prepared for a particular viewing condition. They are related. They are not interchangeable.

Image-generation workflows are judged on visible pixels. Give them a picture with an intentional exposure and view transform, not an unviewed array of linear light values and a prayer. Preserve the EXR separately because it remains the source for grading, relighting, masks and recovery.

This echoes the hierarchy in our ComfyUI control-map guide: authoritative renderer data should remain authoritative. The AI branch may interpret an image. It should not overwrite the evidence that made the image.

What should the handoff file contain?

Export a lossless, display-ready image at the working resolution. PNG or TIFF is a sensible interchange choice when the downstream nodes support it. Use the same approved camera crop as the EXR. Embed or record the intended color profile, and name the view transform in the filename or sidecar notes.

Keep data maps out of color transforms. Depth is not a dark photograph. Normals are not a strange blue wall finish. OpenColorIO explicitly recognizes a data category for non-color channels. Mark and route those files accordingly in applications that support the distinction.

Does denoise strength protect color?

No. It controls how far generation may depart from the image representation the model receives; it does not certify the color pipeline.

ComfyUI's official Stable Image Ultra workflow describes image denoise from 0 to 1, with lower values staying closer to the input and higher values behaving more like text-to-image generation. That is useful behavior to document. It does not guarantee that a low value will preserve HDR range, an embedded profile or a renderer's exact tone curve.

Use denoise to set the permitted visual change, then use comparison tools to inspect the result. Our ComfyUI change-budget method defines what may move before the graph runs. Color needs its own protected fields: white balance, facade value relationships, window highlight detail and shadow readability.

How do you test the round trip?

Make a small reference strip from the input before generation. Include a bright window, a deep interior shadow, a neutral gray material, a saturated accent and a smooth sky or wall gradient. Keep it beside the output, not inside the prompt.

After generation, view both files through the same managed display path. Check the histogram for clipped ends, but look at the building too. Histograms cannot tell you that warm concrete became pink or that two specified gray finishes collapsed into one fashionable greige.

Inspect smooth areas at delivery size and at 200 percent. AI detail and eight-bit processing can produce banding or texture where a calm gradient belonged. Toggle the output against the input with a difference blend. Large intended changes should correspond to the brief. Large unexplained changes are defects, however handsome.

What comes back into the compositor?

Prefer a bounded contribution. If the AI pass repaired planting, isolate the accepted planting and place it over the renderer master using checked edges. If it improved one material, use a trusted mask rather than accepting its new sky, glass and paving as collateral.

The source render can then supply windows, reflections and tonal range that the generative pass weakened. This is the same logic behind exporting a native depth pass before ComfyUI: retain what the scene knows instead of asking pixels to rediscover it.

Do not expect a generated eight-bit PNG to expand back into the original scene-linear range. Converting its profile or placing it in an EXR container changes how numbers are interpreted; it does not resurrect clipped information. Composite the useful visible change over the file that still has the information.

Which color mistakes deserve an immediate stop?

Stop when a generated pass clips a decision-critical window, merges two specified materials, shifts a brand color, reverses a warm-cool lighting relationship or introduces banding into a deliverable gradient. These are not finishing quirks. They change what the image says.

A slight global cast may be correctable if it is consistent and the altered regions remain separable. Local color drift across repeated facade panels is more troublesome because one correction will not fix every instance. Mask it, lower the change strength or give the task back to the compositor.

Reflections deserve special suspicion. They carry both material and light, so a model may “improve” them by inventing a brighter source or a different room. Our reflection repair guide treats geometry, material and lighting checks as separate questions. Color management keeps that diagnosis from beginning with a damaged file.

How should the workflow be recorded?

Save the renderer version, source EXR name, output transform, exposure offset, interchange format, embedded profile, ComfyUI workflow, model, seed and change strength. Record which AI regions were accepted and how they were composited.

Include one untouched input thumbnail and one approved output thumbnail in the issue record. Names and numbers explain the setup; the pair makes an accidental transform or crop change visible before the next revision inherits it.

That sounds fussy until the same camera returns two weeks later with revised cladding. A short record lets the new pass meet the old one without an archaeological dig through screenshots named final-final-3.

The record also prevents a common category mistake. A vivid AI output may differ because the model changed the scene, because the view transform changed the display, or because the profile was ignored. Those failures need different repairs. One undifferentiated “color issue” wastes everybody's morning.

Our take: let the EXR keep the keys

ComfyUI can be a precise post-production instrument when its assignment is precise. Give it an image meant to be seen, a change boundary meant to be obeyed and control maps meant to remain data.

Keep the linear render and passes nearby, boring and intact. They carry the exposure range, scene relationships and revision path that a generated image cannot promise. The AI pass may bring flowers. It does not get the deed.