How to Estimate the Relative Compute Cost of Any ComfyUI Workflow?

[chendind]

Hi everyone,

I'm working on a system where users can run arbitrary ComfyUI workflows, and I need a way to estimate the compute cost (relative or normalized) of any given workflow before actually running it — ideally, to enable dynamic pricing and resource allocation.

Currently, the only method I can think of is:

Run the workflow on a known GPU (e.g., A100), and use the wall-clock execution time as a proxy for compute cost.

This does give a reasonable estimate, but it comes with some significant downsides:

Changing any model, resolution, or setting requires re-running the full workflow to get updated compute cost.

Not scalable when dealing with many user-defined workflows.

I'm wondering:

Has anyone already explored this problem?

Is there a way to predict compute cost without running the workflow, perhaps by statically analyzing the graph structure, node types, model sizes, resolutions, etc.?

Are there known heuristics, profiling tools, or approaches used elsewhere (like in ML compilers or model schedulers) that could apply here?

Any suggestions, references, or even partial ideas would be really appreciated 🙏

Thanks in advance!