Seriously like I am stupid and new to lora training yet it trains easier than an SDXL finetune does. Yes for gooner gens people I know many people are wondering this and it can do it out of the box even. Like out of the box feels like almost as capable as most gooner fine tunes. For people who want to train heres the config im using in WSL. Nice job on the model. Training at 1024 resolution. Efficient as well only using 13gb of VRAM. Also if you see stupid ass settings in my config please tell me. Like I said I am new and stupid lol.

anima.toml

output_dir = '/home/WSL/diffusion-pipe/output/anima'

dataset = 'examples/dataset.toml'

epochs is at the default number since no idea how long it will take to train fully

epochs = 1000
micro_batch_size_per_gpu = 1
pipeline_stages = 2
gradient_accumulation_steps = 4
gradient_clipping = 1.0
warmup_steps = 100

eval_every_n_epochs = 1
eval_before_first_step = true
eval_micro_batch_size_per_gpu = 1
eval_gradient_accumulation_steps = 1

save_every_n_epochs = 2
checkpoint_every_n_minutes = 120
activation_checkpointing = 'unsloth'

May be required for pipeline_stages > 1

reentrant_activation_checkpointing = true
partition_method = 'parameters'
save_dtype = 'bfloat16'
caching_batch_size = 1
steps_per_print = 1

[model]
type = 'anima'
transformer_path = '/home/WSL/diffusion-pipe/anima/diffusion_models/anima-preview.safetensors'
vae_path = '/home/WSL/diffusion-pipe/anima/vae/qwen_image_vae.safetensors'
llm_path = '/home/WSL/diffusion-pipe/anima/text_encoders/qwen_3_06b_base.safetensors'
dtype = 'bfloat16'

Disable llm_adapter training for stability on small datasets.

Set to a small value (e.g. 1e-6) or remove this line to enable it for larger datasets.

llm_adapter_lr = 0

[adapter]
type = 'lora'
rank = 32
dtype = 'bfloat16'

[optimizer]

Anima may need lower LR than other models per the docs.

type = 'adamw_optimi'
lr = 1e-5
betas = [0.9, 0.99]
weight_decay = 0.01
eps = 1e-8

dataset.toml

Resolutions to train on, given as the side length of a square image. You can have multiple sizes here.

!!!WARNING!!!: this might work differently to how you think it does. Images are first grouped to aspect ratio

buckets, then each image is resized to ALL of the areas specified by the resolutions list. This is a way to do

multi-resolution training, i.e. training on multiple total pixel areas at once. Your dataset is effectively duplicated

as many times as the length of this list.

If you just want to use predetermined (width, height, frames) size buckets, see the example cosmos_dataset.toml

file for how you can do that.

resolutions = [1024]

You can give resolutions as (width, height) pairs also. This doesn't do anything different, it's just

another way of specifying the area(s) (i.e. total number of pixels) you want to train on.

resolutions = [[1280, 720]]

Enable aspect ratio bucketing. For the different AR buckets, the final size will be such that

the areas match the resolutions you configured above.

enable_ar_bucket = true

The aspect ratio and frame bucket settings may be specified for each [[directory]] entry as well.

Directory-level settings will override top-level settings.

Min and max aspect ratios, given as width/height ratio.

min_ar = 0.5
max_ar = 2.0

Total number of aspect ratio buckets, evenly spaced (in log space) between min_ar and max_ar.

num_ar_buckets = 7

Can manually specify ar_buckets instead of using the range-style config above.

Each entry can be width/height ratio, or (width, height) pair. But you can't mix them, because of TOML.

ar_buckets = [[512, 512], [448, 576]]

ar_buckets = [1.0, 1.5]

For video training, you need to configure frame buckets (similar to aspect ratio buckets). There will always

be a frame bucket of 1 for images. Videos will be assigned to the longest frame bucket possible, such that the video

is still greater than or equal to the frame bucket length.

But videos are never assigned to the image frame bucket (1); if the video is very short it would just be dropped.

frame_buckets = [1]

If you have >24GB VRAM, or multiple GPUs and use pipeline parallelism, or lower the spatial resolution, you could maybe train with longer frame buckets

frame_buckets = [1, 33, 65, 97]

Shuffle tags before caching, 0 to keep original caption (unless shuffle_tags is set to true). Increases caching time a tiny bit for higher values.

cache_shuffle_num = 10

Delimiter for tags, only used if cache_shuffle_num is not 0. Defaults to ", ".

"tag1, tag2, tag3" has ", " as delimiter and will possibly be shuffled like "tag3, tag1, tag2". "tag1;tag2;tag3" has ";" as delimiter and will possibly be shuffled like "tag2;tag1;tag3".

cache_shuffle_delimiter = ", "

[[directory]]

Path to directory of images/videos, and corresponding caption files. The caption files should match the media file name, but with a .txt extension.

A missing caption file will log a warning, but then just train using an empty caption.

path = '/home/WSL/diffusion-pipe/mydataset'

You can do masked training, where the mask indicates which parts of the image to train on. The masking is done in the loss function. The mask directory should have mask

images with the same names (ignoring the extension) as the training images. E.g. training image 1.jpg could have mask image 1.jpg, 1.png, etc. If a training image doesn't

have a corresponding mask, a warning is printed but training proceeds with no mask for that image. In the mask, white means train on this, black means mask it out. Values

in between black and white become a weight between 0 and 1, i.e. you can use a suitable value of grey for mask weight of 0.5. In actuality, only the R channel is extracted

and converted to the mask weight.

The mask_path can point to any directory containing mask images.

#mask_path = '/home/anon/data/images/grayscale/masks'

How many repeats for 1 epoch. The dataset will act like it is duplicated this many times.

The semantics of this are the same as sd-scripts: num_repeats=1 means one epoch is a single pass over all examples (no duplication).

num_repeats = 1

Example of overriding some settings, and using ar_buckets to directly specify ARs.

ar_buckets = [[448, 576]]

resolutions = [[448, 576]]

frame_buckets = [1]

You can list multiple directories.

[[directory]]

path = '/home/anon/data/images/something_else'

num_repeats = 5