suzushi/miso-diffusion-2.1

Change in model behavior

Trained with brand new dataset this time, it might take some time or more epoch before the model can properly adapt, please be aware that some older prompts might degrade or create some artifact.

Version History

Version	Base Training	Aesthetic Training	Total Epochs
alpha	250K images	0 images	1
beta	160K images	0 images	3
1.0	600k images	0 images	2 + (3 from beta)
1.1	710k images	0 images	5
2.0	1.08M images	0 images	5
2.1	1.32M images	0 images	5

Training Methodology

Training is done on gh200 with 96gb vram, now that prior training shows decent results, I am slightly increasing learning rate.

Training setting: Adafactor with a batchsize of 40, lr_scheduler: cosine SD3.5 Specific setting: enable_scaled_pos_embed = true

pos_emb_random_crop_rate = 0.2

weighting_scheme = "flow" learning_rate = 8e-6

learning_rate_te1 = 5e-6

learning_rate_te2 = 5e-6

Train Clip: true, Train t5xxl: false

Support Me

At the moment training an epoch cost around 130 dollars. If you like my project please consider supporting me: https://ko-fi.com/suzushi2024

Lastly, huge thanks to meg who has been supporting this project, without him this project would not have been possible !