Why is the maximum seed value in Automatic1111 18446744073709519872 if the max seed in manual_seed is 18446744073709551615?

[blobbfobb]

Seeds are processed by PyTorch's manual_seed function which accepts seeds in the range between -9223372036854775808 and +18446744073709551615:

seed (int) – The desired seed. Value must be within the inclusive range [-0x8000_0000_0000_0000, 0xffff_ffff_ffff_ffff]. Otherwise, a RuntimeError is raised. Negative inputs are remapped to positive values with the formula 0xffff_ffff_ffff_ffff + seed.

Source: https://pytorch.org/docs/stable/generated/torch.manual_seed.html

Why is the actual maximum seed value in Automatic1111 lower than that? Higher values than 18446744073709519872 return an error: RuntimeError: Overflow when unpacking long. And where in the code does this limitation happen?

[psykokwak-com]

Take a look here :

stable-diffusion-webui/modules/processing.py

Line 686 in feee37d

return int(random.randrange(4294967294))

[w-e-w]

0xffff_ffff_ffff_ffff - 18446744073709519872 = 31743
this number is very close to the "ENSD 31337" used for reproducing NovelAI images (not sure if this is still valid)
so I'm making the assumption that OP was using ENSD around 31743
SEED + ENSD > 0xffff_ffff_ffff_ffff RuntimeError: Overflow when unpacking long may occur

---

now as psykokwak-com mentioned, the when seed -1 is used we only
Generate random seed between [0, 0xFFFF_FFFF)
https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/feee37d75f1b168768014e4634dcb156ee649c05/modules/processing.py#L686
the reason we limited it to 16bit is because limitation of Gradio (the web interface freamwork)
Gradio treats Numbers as 16bit floats and so will have precision issues with larger values

this limitation can be overcome by using a text element as opposed to a number element and we do have an option to do this --use-textbox-seed
using --use-textbox-seed you'll be able to use the full range of seed

that being said it's just RNG SEED
unless you're planning to generate more then 0xFFFF_FFFF images of the same input (Prompt, Sampler ... every variable the same)
using big values doesn't give you an advantage there's really no practical reason to use long seeds

16bit seeds is more than enough, they are just seeds