Flux becomes slower and more Vram intensive with longer prompts?

[BrechtCorbeel]

Every time I use a long prompt it seems my Vram goes out of whack, anyone else notice this too?

[Melyns]

I also experience this... i ain't no expert but i think it's due to its architecture, more tokens = less iterations per second.

[BrechtCorbeel]

I find it odd, this was never an issue with stable diffusion.

[comfyanonymous]

That's perfectly normal. It's the same with all SD versions but the impact is a lot more significant with DiT models like flux because of how they work.

[BrechtCorbeel]

I never had any problems with this in SD hence the question because I started trying my SD prompts and then the issues started.

Would it be possible to develop something in comfy to estimate Vram use?, because FLUX + memory usage is getting out of hand and unpredictable.

[al-swaiti]

try gguf version of t5 ,, plus run in --lowvram mode https://civitai.com/models/668417/t5gguf

[al-swaiti]

i run bf16 t5 gguf with flux dev bf16 gguf on 12 G gpu >>>> but the equation of time go exponential to the power to the ( size of model / size of gpu ) so i used q8 and q6k ,i created interesting merge here try it and plz write your feedback https://civitai.com/models/682369