This is my first Lora, recommended weight: about 0.8 (SD1.5) / 1.0 (SDXL).

If you have any suggestion on how to improve it, please let me know.

It mainly serve as test ground for different hyperparameters.

TODO: crop out some signature so that it'll be optional.

Update 2025-06-06:

Retrained Flux.1 Dev version with diffusion-pipe + Automagic Optimizer, 8 Rank Lora turns out to be surprisingly good. Now using RTX4080 (better power efficiency & performance)

Update 2025-02-02:

Animagine v4 version

Update 2025-01-28:

Retrained on NoobAI V-pred v1.0, with Prodigy optimizer and cosine annealing.

Update 2024-11-25:

Retrained with t-illunai3 SDXL checkpoint for better prompt adherence, trained with Prodigy optimizer and cosine annealing on a 2080Ti (enabled multires noise)

Update 2024-10-03:

Further trained to 4000 steps

Update 2024-09-26:

Retrained with natural language captioning on 3090Ti with Prodigy optimizer.

Update 2024-09-08:

Flux.1 Dev Lora trained on 3090Ti with SD1.5-style tags. Haven't got time to do natural language captioning.

Update 2024-06-28:

Retrained it on 3090TI, both TE+Unet are trained this time because of huge 24G VRAM. adafactor optimizer works perfectly.

Update 2024-02-06:

Retrained it on Tesla P40 at FP32 precision, it ran for 11.5 hours.

Update 2024-01-27:

Retrained the same dataset on Animagine XL v3.0 base, with gradient checkpointing enabled with my RTX3080, it barely fits into the VRAM.

Update 2024-01-23:

I manually tweaked all prompts in the dataset, now it should be easier to get non-black latex.

Update 2023-12-05:

Retrained on a RTX3080, now at 768px resolution, details improved significantly.