When a ComfyUI result changes after you combine conditioning inputs, keep the workflow structure fixed and, as an author-recommended comparison method, replace one CONDITIONING input at a time with a defined, reproducible control. The control might be a separately prepared conditioning output, a documented neutral prompt or other explicitly chosen input; it must be named in the record rather than assumed to exist in the baseline workflow. Keep the other conditioning connection and the downstream graph unchanged. Compare each result with the baseline and record what changed. This identifies which input change is associated with an output difference in that workflow; it does not prove a universal cause or predict how every model will behave. The cited official documentation supports only the narrower ConditioningZeroOut behavior: it sets pooled_output to zero where that element is present, not that it removes the entire conditioning input.
Define a useful baseline
Choose the output you want to diagnose and save a copy of the exact workflow that produced it. Record the model, relevant node settings, prompt and seed if available, and any other settings you can hold constant. These are controls for your comparison, not a claim that every workflow exposes the same fields.
Label the two conditioning paths clearly, such as A: subject description and B: lighting description. Mark the CONDITIONING sockets where they enter the merge or downstream node. If the result is stochastic or a setting cannot be held constant, note that limitation; otherwise, variation unrelated to your edit can make the comparison harder to interpret.
Isolate one input at a time
- Run or retain the original workflow output as the baseline. Record the visible features you are checking, such as subject pose, background or lighting, rather than only writing “better” or “worse.”
- Make a separate comparison copy. Change only input A: replace that
CONDITIONINGconnection with a separately prepared, explicitly named control input, such as a documented neutral prompt's conditioning output. Do not treat the original A connection as its own replacement control. Keep input B and the downstream graph unchanged. - Compare the new output with the baseline and record both the edit and the observed differences. Avoid changing prompt text, merge settings or unrelated nodes in the same pass.
- Restore A to the baseline state. Prepare or select the documented control for input B, then repeat the comparison for B, again changing no other relevant input.
- If an input is hard to replace cleanly and you choose
ConditioningZeroOut, record that the probe targetspooled_outputelements where present. Do not describe it as deleting, disconnecting or neutralizing every part of that conditioning branch.
ComfyUI documents ConditioningZeroOut as accepting and returning CONDITIONING, with the operation applied to pooled_output when present. See the ConditioningZeroOut node documentation. Since that page describes a specific field operation, it does not establish that other conditioning content or downstream effects disappear. Treat it as a distinct probe, not an equivalent substitute for replacing the whole input.
Keep an evidence-focused record
Use a row per comparison, and distinguish the edit from what you actually saw. The entries below are an original hypothetical example, not a ComfyUI run or a promised model response. “Changed” describes the hypothetical observation only.
| Pass | Input changed | Operation | Hypothetical observation against baseline |
|---|---|---|---|
| Baseline | None | Original A and B connected | Subject left of frame; neutral background |
| A check | A: subject description | Replace A with the separately prepared control conditioning recorded for this pass | Subject placement appears changed; background looks similar |
| B check | B: lighting description | Replace B with the separately prepared control conditioning recorded for this pass | Background appears warmer; subject placement looks similar |
For this hypothetical case, the notes make A the branch to inspect for an association with the placement difference and B the branch to inspect for an association with the background color difference. They do not establish that either description alone caused those features, or that the pattern will hold with different settings or models. If you used ZeroOut, add a separate entry naming it and specifying pooled_output, rather than merging that result into a whole-input replacement row.
Interpret ambiguous comparisons
If replacing a branch produces no obvious difference, report “no clear difference observed under these settings,” not “this condition has no effect.” A feature may be subtle, not visible in the selected output, or entangled with another input. Repeat the comparison only after identifying a concrete uncertainty: for example, whether the same feature changes when A is replaced while B stays fixed.
If both individual comparisons differ from baseline, both changes may be relevant in this particular setup, or the observations may reflect uncontrolled variation. If neither isolates the change, inspect the merge and the downstream graph, then test a clearly defined combined-input comparison. Do not infer a single driver from an output that changes in several ways at once.
Common questions
Should I remove the whole branch or use ZeroOut?
Use whole-input replacement when your question is what happens if that CONDITIONING input is changed to a defined control. Use ZeroOut only when you specifically want to probe the documented pooled_output operation. The two operations are not interchangeable.
What if the branch has no `pooled_output`?
The documentation says ZeroOut changes pooled_output where present. It does not say the node removes other fields when that element is absent. Record the operation’s scope and use a separate, explicit input replacement if you need to compare the whole branch.
Can one comparison tell me which branch controls the image?
It can show which edit is associated with a visible difference in the tested workflow. It cannot establish a general rule about all outputs, models or settings. Keep the graph and comparison conditions fixed, and describe the result at the level you observed.
Sources and next step
This guide interprets public documentation and uses explicitly hypothetical acceptance examples, not application tests. Controls and availability can change; verify the cited manual against your installed version.
For work created in Imagild, carry the input requirements and acceptance checklist into the photo editor. The external software controls discussed above belong to their respective products.