SDXL 1.0 strange / garbled outputs

I know it's early days, but are there any guidelines on SDXL 1.0 training?
These are the settings I used with v0.9 to get perfect outputs (I know, 30k steps is a lot, but I gradually increased to that with experimentation for this large dataset):
`accelerate launch sdxl_train_network.py --network_module=networks.lora --pretrained_model_name_or_path=E:/Downloads/stable-diffusion-webui/models/Stable-diffusion/sd_xl_base_0.9.safetensors --dataset_config=model.toml --output_dir=outputs --output_name=model --save_model_as=safetensors --prior_loss_weight=1.0 --max_train_steps=30000 --optimizer_type="Adafactor" --xformers --mixed_precision="bf16" --cache_latents --gradient_checkpointing --lr_scheduler="constant_with_warmup" --learning_rate=4e-7 --no_half_vae --network_dim=128 --network_args "conv_dim=128"`

But now with v1.0, I get results like this. Lots of deformities, random splotches of stuff, repeating details, overall low quality, etc.

I would say it's just being overbaked, except that it's not following the prompt well. In fact, it's following it less.
I get results like this when using a v0.9 LoRa with V1.0 base too.
Would be nice to know if it's just me before training again because if it's just being overbaked, it'll take several days to find the right parameters again, and if not, I'll have wasted all that time for nothing

Start with the reported command for sdxl_train_network.py and the model.toml dataset configuration, comparing the v0.9 and v1.0 settings and outputs shown in the issue. Determine whether the garbled results come from a v1.0 training incompatibility or configuration difference; done means identifying reproducible guidance or a confirmed defect.

Written by the indexing model from the issue text.