⚠️ WALL OF TEXT INCOMING! ⚠️

Thanks glad you like this!
My recommended settings are the default settings from here: https://github.com/citronlegacy/citron-anima-lora-trainer-ui (I listed all the settings at the bottom of this comment).

If you don't have much experience making Loras I recommend using my app because its focus is to be user friendly. (It only takes about 6GB VRAM)

--

As for dataset size I've made loras with just 1 image on Anima-Preview3,

https://civitai.red/models/2538818/chuni-pose-or-chunibyo-pose-concept-anima

https://civitai.red/models/2536257/virgin-islands-meme-concept-anima

But these aren't very flexible 🤷‍♂️

--

Anima Base v1 is less forgiving for small datasets but does work. For styles I prefer larger datasets.
This CuteRetroStyle_ANIMAv1 lora was made with 265 images and 3 repeats.

This https://civitai.com/models/185413/ lora was made with 794 images and 4 repeats.
This https://civitai.red/models/1995901/ lora was made with 107 images and 3 repeats.
The https://civitai.red/models/178169/ lora was made with 541 images and 1 repeats.
In general you can see the training settings for any lora but using this tool https://citronlegacy.github.io/lora-metadata-viewer/
--
For Anima-base-v1 or Preview3 use these same settings but change the Dim and Alpha.
I like 30 for Dim and Alpha on Anima-base-v1. You can use 30 for Preview3 or 20 Dim and Alpha is good.
Number of repeats depends on number of images. Also Anima will resist some styles so sometimes you need more repeats.

--

You can mess around with settings to get different results but 99% of the time I only change Dim, Alpha, and repeats.
All the other settings can stay the same

network_dim = 20

network_alpha = 20

learning_rate = 0.0001

max_train_epochs = 10

resolution = 768

repeats = 10

caption_dropout = 0.1

optimizer_type = "AdamW8bit"

lr_scheduler = "cosine_with_restarts"

lr_scheduler_num_cycles = 1

lr_warmup_steps = 100

train_batch_size = 1

gradient_accumulation_steps = 1

max_grad_norm = 1.0

save_every_n_epochs = 1

save_last_n_epochs = 4

mixed_precision = "bf16"

gradient_checkpointing = True

seed = 42

noise_offset = 0.03

multires_noise_discount = 0.3

timestep_sampling = "sigmoid"

discrete_flow_shift = 1.0

cache_latents = True

cache_text_encoder_outputs = True

vae_chunk_size = 64

vae_disable_cache = True

num_cpu_threads_per_process = 1