/vt/ - Virtual Youtubers

Anonymous Tue, Oct 10, 2023 11:10:27 No.60218039 [View]
File: 265 KB, 922x510, Screenshot vae.png [View same] [iqdb] [saucenao] [google]

>>60217405
description says he used
>about 22k labeled images to train about 100 epochs on a cosine scheduler with a learning rate of 5e-6

HLL usually uses about 100-200k images and 10-15 epochs. If SDXL requires significantly longer training time, I see no reason to touch it.
SDXL will never be good anyway, people who made it did everything wrong. Its VAE is shit (4 channel on picrel is SDXL), 2 text encoders are shit, unet and idea of using refiner is also shit

Advanced search
Text to find
Subject [?]Search by post subject. Leave empty for any.
Username [?]Search for user name. Leave empty for any user name.
Tripcode [?]Search for tripcode. Leave empty for any.
Email [?]Search by email. Leave empty for any.
Filename [?]Search by image filename. Leave empty for any.
From Date [?]Enter what date to start searching from. Format is YYYY-MM-DD
To Date [?]Enter what date to start searching until. Format is YYYY-MM-DD
Image hash
Search in	All Posts OPs Only
Deleted posts	Show all posts Show only deleted posts Only show non-deleted posts
Internal posts	Show all posts Show only internal posts Show only archived posts
Order	New posts first Old posts first
Capcode	All Posts Only by Users Only by Mods Only by Admins Only by Developers
Results	Posts Threads
Action	[ Simple ]

Navigation
View posts	[+24]	[+48]	[+96]

/vt/ - Virtual Youtubers

Search: