circlestone-labs

/

Anima

like 949

Follow

CircleStone Labs 489

You deffinitely shouldn't be using noise offset for this model, that's the first and only thing i can pretty confidently say is "wrong."

that could 100% lead to weird colors and grainy stuff happening where it shouldn't. try without.

Try default config.

Alpha should be half of dim if you don't want to overtrain quickly.
network_dim=8
network_alpha=4

learning_rate=2e-4

Remove these fancy stuff that's not even required:
prior_loss_weight = 1.0
edm2_loss_weighting = false
timestep_sampling = "sigma"
discrete_flow_shift = 3.0
weighting_scheme = "uniform"
debiased_estimation_loss = true
noise_offset = 0.1

Also remove Loraplus

How are you even able to train it, it's constantly giving me a 'CUDA out of memory' error ;_;

64gb of ram and 24gb of vram

How are you even able to train it, it's constantly giving me a 'CUDA out of memory' error ;_;

What are your specs?

How are you even able to train it, it's constantly giving me a 'CUDA out of memory' error ;_;

What are your specs?

it's the same Colab as OP so it's a T4 but he's seemingly able to train it just fine.
EDIT: I'm silly and didn't cache latents earlier. It works now! Hopefully I don't have OP's issues with it.

Hey guys, thanks for the help!

Huge facepalm from me… turns out it wasn’t a training problem at all.

I was testing with the “image → VAE” node and had lowered the noise in the KSampler. After I was done, I forgot to set the noise back to 1.0. So I wasted days retraining when the green tint was just a dumb workflow mistake. Classic retarded noobie error.

That said, your advice still helped a lot! Removing noise_offset, setting alpha to half the dim, and stripping out all the extra fancy stuff made the LoRA train quite a bit cleaner.

Here are the new results with the fixed workflow (same prompt/seed):

And here’s my current (simpler) TOML:

[[subsets]]
caption_extension = ".txt"
image_dir = "/content/drive/MyDrive/Loras/[ARTISTSTYLE]/dataset"
name = "subset 1"
num_repeats = 1
random_crop_padding_percent = 0.0

[train_mode]
train_mode = "lora"

[general_args.args]
persistent_data_loader_workers = true
pretrained_model_name_or_path = ""
gradient_checkpointing = true
seed = 23
max_data_loader_n_workers = 1
max_token_length = 225
sdpa = true
max_train_epochs = 20
cache_latents = true
vae_batch_size = 1
full_bf16 = true
mixed_precision = "bf16"

[general_args.dataset_args]
resolution = 1024
batch_size = 1

[network_args.args]
network_dim = 16
network_alpha = 8.0
min_timestep = 0
max_timestep = 1000

[optimizer_args.args]
optimizer_type = "AdamW8bit"
lr_scheduler = "cosine"
loss_type = "l2"
learning_rate = 0.0002
warmup_ratio = 0.1
max_grad_norm = 1.0

[saving_args.args]
output_dir = "/content/drive/MyDrive/Loras/[ARTISTSTYLE]/output"
save_precision = "bf16"
save_model_as = "safetensors"
save_every_n_epochs = 1
save_last_n_epochs_state = 1
save_state = true
output_name = "[ARTISTSTYLE]"

[logging_args.args]
log_prefix_mode = "disabled"
run_name_mode = "default"

[anima_args.args]
pretrained_model_name_or_path = "/content/drive/MyDrive/Downloaded_models/anima-preview2.safetensors"
qwen3 = "/content/drive/MyDrive/Downloaded_text_encoders/qwen_3_06b_base.safetensors"
vae = "/content/drive/MyDrive/Downloaded_VAEs/qwen_image_vae.safetensors"
qwen3_max_token_length = 512
t5_max_token_length = 512
timestep_sampling = "sigmoid"
sigmoid_scale = 1.0
discrete_flow_shift = 3.0

[edm_loss_args.args]
edm2_loss_weighting = false

[bucket_args.dataset_args]
enable_bucket = true
min_bucket_reso = 512
bucket_reso_steps = 64
max_bucket_reso = 1536

Also @brokencontroller ,

Quick tip since you're on the free T4 and can only train for 1.5–4 hours at a time:
Make sure you keep save_state = true and save_last_n_epochs_state = 1. That way you can resume training from where you left off instead of starting over every time Colab disconnects.

Learned my lesson, always double-check the workflow before (re)training! Thanks again :