When a ControlNet result misses the intended structure, inspect the preprocessed image actually connected to ControlNetApply before changing strength. The original image is not necessarily the control signal the node receives. First decide whether that signal shows the features you want to preserve; only then compare strengths while keeping the prompt, model, and other generation settings fixed. This separates an input-content problem from a constraint that may be too strong.
Define what the control image must preserve
Choose two or three visible criteria before opening the graph. For a line-art composition, these might be the outer silhouette, the position of a window, and the direction of a roofline. Also note details the control does not need to preserve, such as paper texture. This prevents “looks wrong” from turning into an unstructured strength adjustment.
ComfyUI’s ControlNetApply documentation says the input image must be preprocessed. It describes image as an image processed by a preprocessor, and identifies strength as the amount of network adjustment: lower values permit more model freedom, while higher values impose stricter constraints. The documentation gives a range and a suggested interval, but that is not a guarantee that a particular value suits every model or image. See the ControlNetApply node documentation.
Inspect the signal that reaches the node
- Find the image output connected to
ControlNetApply’simageinput. Identify the preprocessing node or path upstream of that connection. - Display or otherwise inspect that exact output. Do not substitute the source photo, an earlier intermediate, or what you expected the preprocessor to produce.
- Compare the displayed output against your criteria. Ask whether the important outlines or spatial relationships are visible, missing, broken, or overwhelmed by unrelated marks.
- Record the control model, preprocessor and its settings, prompt, generation settings, and current strength. If the preprocessed output does not express the target structure, adjust the input or preprocessing setup first; then inspect the resulting output again.
The ControlNetApply input reference distinguishes the image input from the strength and guidance timing inputs. This makes the diagnostic order practical: establish what image is being applied before judging how forcefully it is applied.
Compare strength only after the input passes
Keep the inspected control image, prompt, model, seed if available, sampler, and other generation conditions unchanged. Compare a small set of supported strength values, changing only strength. Record results even when they fail. The purpose is not to find a universal number; it is to see whether the same suitable signal becomes overly restrictive as its influence increases.
Treat start_percent and end_percent as separate controls. The documentation describes them as when guidance begins and ends during diffusion. Leave them fixed during a strength comparison so that a timing change does not muddy the evidence. If you later investigate timing, make it a separate comparison.
Hypothetical acceptance record
The following is an invented line-art example, not a reported test. A creator wants a doorway position and roof outline retained in a generated house scene. The actual preprocessor output appears to preserve the roof but omit the doorway. The output therefore fails the input check before strength is considered.
| Trial | Actual control image | Fixed conditions | Strength | Hypothetical observation | Decision |
|---|---|---|---|---|---|
| Input check | Doorway absent; roof visible | Record model, prompt, and settings | Current value | Target structure is incomplete in the signal | Revise or replace the input/preprocessing path, then inspect again |
| A | Same inspected image | Same model, prompt, seed, and settings | 0.6 | Roof is present; doorway remains absent | Do not attribute the missing doorway to strength alone |
| B | Same inspected image | Same model, prompt, seed, and settings | 1.0 | Roof appears more restrictive; doorway still absent | Possible excessive constraint for this trial, but input mismatch remains |
These numbers are example entries, not recommendations or predicted outcomes. A stronger setting cannot be assumed to restore a feature missing from the control signal. Conversely, an adequate control image does not prove that strength is the only cause of an unsatisfactory result.
What to do when the diagnosis is unclear
If the displayed control output is blank, broken, or unrelated to the target, investigate the preprocessing path and confirm the image connected to the node before tuning strength. The documentation notes that ComfyUI’s initial nodes do not include preprocessors or ControlNet models and points users to separately installed preprocessors and corresponding models. It does not specify which preprocessor is appropriate for every task, nor promise that a particular combination will retain a chosen detail.
If the input passes your structural check but results become unusually constrained at higher strengths, compare lower supported values with all other conditions fixed. If the result still misses the brief, record the remaining uncertainty rather than claiming the input or strength is proven responsible. Avoid changing strength, prompt, timing, and preprocessing together: that removes the ability to identify which change mattered.
Common questions
#### Should I judge the original image or the preprocessor output?
Judge the exact image connected to ControlNetApply’s image input. The original image can help explain the transformation, but it does not establish what ControlNet receives.
#### Are the documented strength values a recipe? No. The documentation describes a range, a suggested interval, and the relative effect of lower and higher strength. It does not guarantee a result for a specific model, preprocessor, or composition.
#### What if the control image looks right but the generation still does not? Keep the control image and other conditions fixed, then compare strength. If the discrepancy remains, report what was held constant and inspect other factors separately; the documentation does not establish that strength alone explains every failure.
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.