ComfyUI parameter control: a reproducible test method
A useful parameter comparison changes one setting while keeping the workflow, model files and inputs fixed. This guide explains how to diagnose the result instead of treating a preferred seed or sampler as a universal improvement.
Start with a known graph
Use the official ComfyUI text-to-image example before adding extensions. It connects a checkpoint, text conditioning and a latent image to sampling, decoding and image saving. Load the provided workflow rather than copying a partial graph from an article. Confirm that all model selectors resolve to files you actually have.
Save that working graph as your baseline. Record the ComfyUI version, model filename and hash, custom-node revisions, accelerator and driver. Save the input image when testing image-to-image. A screenshot is useful for orientation, but it does not preserve every connection or hidden setting needed to repeat a run.
Separate the controls
The KSampler reference distinguishes seed, steps, guidance scale, sampler, scheduler and denoise. Seed controls initial randomness; zero is a valid value, not a diagnosis for black output. Guidance affects conditioning strength, while sampler and scheduler describe different parts of the sampling process. Lower denoise can preserve more of an existing image in an appropriate image-to-image graph.
Do not import a preferred guidance range from a different model family. Start with that model's maintained example. If the graph has a separate guidance node, inspect its role rather than assuming that every field labelled guidance is interchangeable. Record the value and the node that owns it.
Run a controlled comparison
- Choose one visible goal, such as keeping an object's outline while changing its material. Write the acceptance criterion before viewing the outputs.
- Run the baseline with its seed fixed. Check whether the interface increments or randomises the seed after generation; the displayed value must be the one used for the saved result.
- Duplicate the baseline and change one control only. Keep resolution, prompt, batch size and model files unchanged.
- Save both outputs with their workflows and label the changed parameter. Compare the same crop at the same display scale.
- Repeat the comparison with several additional fixed seeds. Keep failures as well as preferred images; a single attractive sample is not evidence of a generally better setting.
If you are testing speed rather than appearance, separate the first run from subsequent runs. Record your timing boundaries: queue waiting, model loading, sampling, decoding and saving are different costs. Do not label the total browser waiting time as sampler speed.
Keep reproduction claims bounded
A fixed seed is a useful control, not a promise of identical output on every computer. PyTorch's reproducibility guidance explains that releases, platforms and nondeterministic operations can affect repeatability. Compare within the same recorded environment before drawing conclusions about a parameter.
When sharing a result, include the workflow, exact inputs, environment notes and an explanation of what changed. State whether the comparison was visually judged or scored with a named metric. An image that you prefer is a valid editorial choice; it should not be presented as a measured accuracy gain.
Using this method with Promptus
If you start a workflow through Promptus, first identify the actual ComfyUI backend and graph it runs. Keep the application version alongside the backend version in your notes. A simplified interface may expose only some graph inputs; inspect the exported workflow before comparing it with a graph launched directly in ComfyUI.
This article does not assert that every Promptus version exposes the same controls or export commands. If an input is hidden, document that limitation rather than inventing a mapping. Use the workflow library to choose an inspectable starting point and the measurement protocol when memory or timing becomes the question.