Workflow doesnt seem to be using correct input image after consecutive runs

[Lalimec]

https://replicate.com/p/va6v99b9qxrmc0cmpza88bsa98

As you can observe from this prediction input images is not being passed to workflow somehow. Result should be an inpainted version of the image with reverse alpha. But input and output doesnt align, it uses some previous input. I guess it can be some kind of cache issue since it reverts back to normal if i force reset cache. I might be completely wrong but i think i am sending the correct file lol.

Here is a correct prediction. https://replicate.com/p/9jar316q59rme0cmpzbscb6mb8 Only thing i did was sending the api request a couple more times.

[fofr]

These have been removed now, can you share again?
Do your input files happen to have the same filename?

[Lalimec]

Yeah sorry about that, i forgot replicate removes them after an hour lol. My input names are unique, i use time stamps in the name. But since we use "input.webp" as the file name maybe comfy thinks they are the same image. I dont know if does or check image hash or sth, if there is even such a thing.

Here are some new predictions run on replicate directly, without force_reset_cache:

https://replicate.com/p/cd70ygyarnrmc0cmqjrr6gq5zm
https://replicate.com/p/6ypayh0xd1rme0cmqjs84asw5m

As you can see, second one still uses the input from the previous one, even though i ve passed a new image with a unique name. I guess i will update the input image field directly with the url, i was sending them with via api.

[fofr]

I'm still debugging what the issue is here. But for now, if you pass in the input file URL in the workflow directly it'll work each time.

[Lalimec]

Yeah will do that thank you.