Lowering denoise in ComfyUI image-to-image does not necessarily lower the number of sampling steps: with a positive denoise value, the documented schedule construction can still perform the full number of steps set in KSampler. The setting also determines which part of a longer schedule is used. To sample fewer steps at the density of a full run, calculate a lower steps value as full-run steps multiplied by denoise, then check the expected range and evaluations rather than judging by the image’s appearance.
Separate the two controls
In the documented image-to-image workflow, steps is not, by itself, a count of the portion of the schedule selected. ComfyUI builds a schedule for steps / denoise, rounded down, then keeps its last steps + 1 sigma values. The resulting sampling run uses the retained interval. This is why lowering denoise while leaving steps unchanged can retain the same number of sampling steps while selecting a lower-noise part of a longer schedule.
The documentation gives a specific rule and two examples, not a guarantee for every model, scheduler, or custom node. Treat the examples below as checks against that documented behavior, not as proof that every workflow uses identical internals. See the ComfyUI image-to-image explanation of the schedule construction and its step-density examples.
Calculate the range before changing the workflow
Use this sequence when you want to reduce work while preserving the step density of a comparison run:
- Choose the step count of the complete run you are using for comparison in the same workflow, such as
20. This is a chosen reference, not a fixed value shared by all models or schedulers; the official20 / 0.5example compares against a 20-step run. - Record the
denoisevalue inKSampler, such as0.5. - For positive
denoise, calculate the schedule length assteps / denoise, rounded down. For20 / 0.5, the documented schedule length is40. - Decide whether to keep the current step count over the longer schedule or use the comparison run's step density for only part of its range. For the latter, calculate
comparison-run steps × denoise:20 × 0.5 = 10. - Check the count and interval against the documented examples below. Keep the checkpoint, input, and other workflow settings unchanged while comparing, so the step-count question is not mixed with other edits.
Do not enter steps / denoise as the new KSampler steps value when your goal is fewer evaluations. That calculation describes the longer schedule from which the retained sigmas are selected; it does not describe the shorter setting in the documentation’s full-density example.
Worked examples and worksheet
The table contains the two step-density examples stated in the official image-to-image documentation. “Range” describes the schedule portion, not a promise about image quality. “Expected sampling steps” follows those examples.
| Full-run reference | Denoise | KSampler steps | Expected sampling steps | Schedule range |
|---|---|---|---|---|
| 20 | 0.5 | 20 | 20 | Low-noise half of a 40-step schedule |
| 20 | 0.5 | 10 | 10 | Same sigmas as the second half of a 20-step run |
For the first row, 20 / 0.5 = 40; retaining the last 20 + 1 sigmas gives the documented low-noise half while the run still performs 20 steps. For the second row, set steps to 20 × 0.5 = 10. The documentation says this selects exactly the sigmas used for the second half of a 20-step run and uses half the model evaluations. Use these rows to check the documented arithmetic and interval, not to predict how attractive either output will look.
Author-suggested worksheet (assumption: the workflow follows the documented image-to-image schedule construction). Record the comparison run's step count, denoise, the KSampler steps setting, the expected sampling-step count, and the schedule interval. Check the first row against 20 / 0.5 = 40 and 20 sampling steps; check the second against 20 × 0.5 = 10, the second-half sigma interval, and half the model evaluations. This worksheet is a way to organize the documented examples, not an official test or a guarantee for other configurations.
Diagnose a mismatch
If your count does not match the table, first verify that you are reading the KSampler steps and denoise fields in the workflow that actually ran. Then confirm that the workflow follows the documented image-to-image setup and that no custom node or alternate sampling path changes how the schedule is built. A progress display or elapsed time is not a direct count of model evaluations; use the workflow settings and the documented calculation for this check.
If you need a precise count for a different scheduler, model, or custom workflow, consult documentation for that configuration or inspect its own execution details. The cited image-to-image page does not establish how every sampler handles rounding, boundary sigmas, custom schedules, or evaluation accounting. Do not extend its examples into guarantees for those cases. Also, fewer model evaluations do not establish better or worse image quality: assess the generated result separately against the visual changes you wanted.
Common questions
Will setting denoise to 0 make a one-step run?
No. The documented page says denoise 0 stops sampling entirely. It is not a way to request one step; use a positive value and an appropriate steps count if you want sampling to occur.
Why does `steps 20, denoise 0.5` still show 20 steps?
Because the documented construction retains steps + 1 sigma values from the longer schedule. The example explicitly says this combination performs 20 sampling steps over the low-noise half of a 40-step schedule.
Does the 10-step setting guarantee the same image as the 20-step setting?
No. The cited page supports the sigma-interval and model-evaluation comparison, not identical outputs or a particular visual result. Evaluate the actual output against your goal.
How can I compare outputs without confusing the step-count check?
As an author-suggested comparison, keep the checkpoint, input image, and other workflow settings fixed, then record the settings, expected evaluations, and output separately for each run. This is a proposed way to isolate the comparison, not an experiment reported by the official documentation.
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.