Please Explain the change? @victor , @julien-c , @hysts , @cbensimon ,
I am as Novice as they come. I rely heavily on the regen rate, for creating images and "fix" the elements I dont want.

Because AI is learning as much as we are and dont get it quite right, I rely on redo, change prompt, redo, change weight, redo, change CFG, redo... redo... sometimes I would do 9 attempts (each costing 90sec) to get a result close enough and still I have to use other software to smooth over and enhance or correct aspects of the results.

**I joined HF as a free user because I could not afford to pay the fees other sites asked.
**I joined HF as a free user because I could benefit from smart programmers models and code to do the things I can not. ( including training of such lora and models I dont have the hardware to do so or else I prob would have, If I learn how.)
**I upgraded to PRO because I needed more "chances" to fix errors in my prompts to get the results I needed.

So please, dont become like many out there that only care about money and not the users.
Yes, I understand AI, is on the rise as the next best thing, but we dont need another Elon Musk money whore.

I sincerely hope this new direction HF is taking would not decrease or ruin, or limit me in such a way that it is not worth paying a subscription.

@cbensimon . @John6666 ,@julien-c ,@victor ,@hysts ,
Thank you. I notice the Quota is now capped at 50min, It does help with the tasks I am doing.

Has the (regen/cooldown) feature been removed completely?
It used to be something like 60s ZeroGPU recovered every 30min ( cant remember exactly).

I dont want to be un-thankful but it was a good feature to have. Even if you slow it down a little ie. (30s regen every 30min) It does allow us users to continue our work just at a slower pace instead of being locked out for 24 hours waiting for the cap to reset.

Having to wait for the reset may also cause some spiked usages at certain times kinda like "rush hour traffic".
I have noticed that some generation seem to fail in such "peak time" cases, and we dont get the time used of the failed ZeroGPU task refunded.

ie, I use 90s for a render, and if it fails without producing an image I lose the 90s and I have to attempt another render spending another 90s. So in effect one render would cost me double.
Not to mention "If" the model miss interpret the prompt and provide a garbage image.
Flux.dev-1. has a CLIP limit and can only handle sequences up to 77 tokens, anything more than that gets truncated leading to inaccuracy's or even ignored details and missing features from the render.

Thank you for understanding,
Any advice to overcome the latter part of my message to increase accuracy and efficiency is most welcome.
I am a total novice at coding so would love and appreciate the help.