FLUX is fast but not with cog-comfyui?

[WesleyKapow]

Trying to understand the overhead of cog-comfyui vs cog-flux and if it is possible to get speeds closer to cog-flux.

With cog-flux, I can generate a flux-dev image in about 3s.

With cog-comfyui, using the fp8 example, it takes about 35-40s to generate an image.

I would hope that caching would resolve any possible loading of models overhead. I assume there's some overhead with running comfy. But it also seems like image generation takes longer too. Are there optimizations from flux is fast and open source that cog-comfyui isn't taking advantage of? Is it possible to drive the timing closer to cog-flux?

Thanks for your help here!!

[jordancoult]

I believe the backend is open source for the flux models on Replicate, so you can see the optimizations they've made for speed and compare them to the Flux node for ComfyUI. It would be great to have those optimizations added as options to the ComfyUI node.