🆕 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<ul><li>Trained with <code>diffusion-pipe</code> on a RunPod A40. Cost: 5$</li><li>Used the following configs:</li></ul><code>config.toml</code> <a target="_blank" rel="ugc" href="https://pastebin.com/YUGnKEPC">https://pastebin.com/YUGnKEPC</a><code>dataset.toml</code> (key settings):<pre><code>resolutions = [1024]
enable_ar_bucket = true
min_ar = 0.5
max_ar = 2.0
num_ar_buckets = 9

[[directory]]
path = '/workspace/dataset/atj'
</code></pre>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: 👉 <a target="_blank" rel="ugc" href="https://discord.gg/PSa9EMEMCe">https://discord.gg/PSa9EMEMCe</a>Lewd stuff currently in process...

Files with SHA256: f3e6c89a21f7d6e6a663ee2e31aff03b62d63c66ea0d3fd45e96c5a3820b3b79