Trigger Word: time magazine
Suggested resolution: 768*1024
Like what I'm up to? Feel free to treat me a Ko-fi. Cheers!
Description
FAQ
Comments (14)
Thx!
May I ask about the training set? Is it just the covers or have the covers been specially edited?
It is just the covers.👍
@NextMeal .. cool. I just downloaded all the covers of DER SPIEGEL (german newspaper) from 2017 :)
@m29498 btw you can consider filter out the images like special edition, as they probably have different format. And I used the logo and rectangular frames in reg images to emphasis them (but I am not sure how much difference that would make, I didn't run many tests on this one).
@NextMeal thx - in 2017 all covers have the same logo/claim/construction and scaled 1024x768 (perfect without cropping).
The only problem I have is that I only have one kohya command, where the LoRa's always become 1.7 GB. That's ok for me, but unfortunate for sharing. Do you know the command line command that leads to your 40 MB LoRa? That would be very, very nice if I could get it.
command line (1,7 GB LoRa) i use is:
accelerate launch --num_cpu_threads_per_process=2 "./sdxl_train_network.py" --enable_bucket --pretrained_model_name_or_path="[path]/sd_xl_base_1.0.safetensors" --train_data_dir="[path]/image" --reg_data_dir="[path]/reg" --resolution="1024,1024" --output_dir="[path]/model" --logging_dir="[path]/log" --network_alpha="1" --save_model_as=safetensors --network_module=networks.lora --text_encoder_lr=0.0004 --unet_lr=0.0004 --network_dim=256 --output_name="[LoRaName]" --lr_scheduler_num_cycles="10" --no_half_vae --learning_rate="0.0004" --lr_scheduler="constant" --train_batch_size="1" --max_train_steps="2048" --save_every_n_epochs="1" --mixed_precision="bf16" --save_precision="bf16" --cache_latents --cache_latents_to_disk --optimizer_type="Adafactor" --optimizer_args scale_parameter=False relative_step=False warmup_init=False --max_data_loader_n_workers="0" --bucket_reso_steps=64 --gradient_checkpointing --xformers --bucket_no_upscale
my first donation :)
@m29498 ahhhh, thank you so much!!! I don't speak very fluent english so I don't know how to express how I feel. Thank you so much!!😭😭
as for your issue, I believe you set your network rank too high, as SDXL has more parameters, I see people having better results when using lower rank than in SD15. For this one, I used network dim = 8.
@NextMeal i will try thx!
@NextMeal it works :)
https://civitai.com/models/161415?
@m29498 that was fast😂 great that it worked!💪
Superman smoking a giant banana...?
\:p
Thank you for the LoRA! Do you find that this makes a big difference?
I feel that SDXL models already have a good understanding of various magazine covers. For example (NSFW) I did these: https://civitai.com/posts/653710
People, Maxim, and Esquire might need more help than Time, from that admittedly small sample set 😂. Looks like you get more of the classic red border on Time magazine when using this LoRA, at least!
SDXL does have a quite extensive knowledge, but still can be not quite accurate. You will get extra letter or misplaced letter more than usual even for popular magazines. I think it just like it can draw celebrities in a certain level of likeness, but still need LoRAs if want to draw them more correctly. As for how much difference that make, that’s for everyone to judge😉
@NextMeal Makes sense. Thanks for the reply and for the LoRA!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.