Ume - Sky
Hello this in my second LoRA.
This time I used more images for creation (30). I tried to give it a “painted” look that I really like.
It was created from images I myself published on Deviant-Art : UmeAiRT - Hobbyist, Digital Artist | DeviantArt
Using too large a weight can result in poor image quality.
Light version is trained on Civitai with only 3,000 buzz.
Don't hesitate to send me feedback or suggestions!
Description
Anima version
FAQ
Comments (5)
how many images did you use to train this lora for Anima? and what were your settings?
40 images,
My exact settings :
[training_arguments]
output_name = "UmeSky_Anima"
save_model_as = "safetensors"
max_train_epochs = 30
save_every_n_epochs = 3
sample_every_n_epochs = 3
learning_rate = 0.0001
text_encoder_lr = 0
optimizer_type = "AdamW8bit"
optimizer_args = [ "weight_decay=0.01" ]
lr_scheduler = "cosine"
lr_warmup_steps = 130
mixed_precision = "bf16"
save_precision = "bf16"
max_data_loader_n_workers = 4
gradient_accumulation_steps = 1
max_grad_norm = 1
gradient_checkpointing = true
flash_attn = false
torch_compile = false
lowram = false
blocks_to_swap = 0
persistent_data_loader_workers = true
seed = 42
cache_latents_to_disk = true
vae_batch_size = 1
cache_text_encoder_outputs_to_disk = false
multigpu_mode = "ddp"
deepspeed = false
use_cuda_direct = false
ddp_gradient_as_bucket_view = false
ddp_static_graph = false
step_profile = false
profile_microbatch = false
[network_arguments]
network_module = "networks.lora_anima"
network_dim = 64
network_alpha = 32
network_train_unet_only = true
auto_resume_last_state = false
[anima_arguments]
timestep_sample_method = "logit_normal"
discrete_flow_shift = 3
weighting_scheme = "logit_normal"
@UmeAiRT thanks, I am going to give this a try. I think I tried similar settings in my lora https://civitai.red/models/2704205/dreamlike-nostalgia-anima but I don't think it captured enough style. I had like ~40 images, I trained for 40 epochs, lr 0.00002, gradient_accu steps 4, also, i used 32, 16 for dim,alpha.
Which tool did you use for this? And what about captions? On some loras I used only trigger word as caption but they seem to be a bit rigid.
@Wheeman My V1 use only 32, 16 for dim,alpha too and result is light, i use 64, 32 for better result? And my caption are tags with my triger word as first. And for more "strengh" i use Keep Tokens 1, Caption Dropout Rate 0.1 and Tag Dropout Rate 0.1
@UmeAiRT going to give it try and share results, thanks!