🆕 New AT-J LoRA

Took HiDream for a test drive and a new AT-J LoRa was the result.

Very smooth learning process overall. Images include prompts and the ComfyUI workflow, as they should.

Is HiDream good?

HiDream is showing real promise as a future base model. Training runs are super smooth, if a bit slow or expensive. Probably both. But that’s likely just a temporary bottleneck until the optimizer nerds finish their work and I can train LoRAs on my toaster.

It’s probably too good of a model. It’ll happily pick up and reflect whatever shit exist in your dataset. Garbage in, garbage out, just really, really nicely rendered garbage.

Training details

• Trained with diffusion-pipe on a RunPod A40. Cost: 5$

• Used the following configs:

config.toml
https://pastebin.com/YUGnKEPC

dataset.toml (key settings):

resolutions = [1024]
enable_ar_bucket = true
min_ar = 0.5
max_ar = 2.0
num_ar_buckets = 9

[[directory]]
path = '/workspace/dataset/atj'

Tried local training on a 4090, but even with block swapping enabled, diffusion-pipe decided to crash and burn. If you get it running locally, let me know how by joining the worst Discord community ever. We have a climate change denier who is quite the funny guy:
👉 https://discord.gg/PSa9EMEMCe

Lewd stuff currently in process...