This was a test run on the original Anima_preview model, similar dataset and training params to my last release. It should still work with anima_base-1.0 but there will be some concept loss until I re-train this on anima_base-1.0

Training settings (lazy posting training most of my kohya config)

export epocs=5
export learning_rate=8e-4 # yes this works fine with scheduler free adamw 5 months in

export te_learning_rate=0

export train_batch_size=4

export gradient_accumulation_steps=32

export mixed_precision="bf16"

export save_precision="bf16"

export save_every_n_steps=200

export sample_every_n_steps=100

export num_cpu_threads_per_process=12

export max_resolution="768,768"

export max_resolution_bucket=1024

export bucket_reso_steps=32

export LR_SCHEDULER="linear"

export network_dim=32

export alpha=$(($network_dim/2))

export conv_dim=$(($network_dim/2))

export conv_alpha=$(($alpha/2))

export tokens=525 # I just left this as is, not even sure if its used in anima

export tag_dropout=0.1

export caption_dropout=0.1

export dropout=0.1

export WARMUP_STEPS=300 #$((${steps} * 5/100))

./venv/bin/accelerate launch --num_cpu_threads_per_process $num_cpu_threads_per_process --num_processes 1 anima_train_network.py \

    --pretrained_model_name_or_path="${pretrained_model_name_or_path}" \

    --in_json $metadata_dir"/meta_cap.json" \

    --train_data_dir="../../datasets/${raw_img_folder}" \

    --output_dir="./output/${short_name}" \

    --resolution=$max_resolution \

    --train_batch_size=$train_batch_size \

    --gradient_accumulation_steps=$gradient_accumulation_steps \

    --learning_rate="${learning_rate}" \

    --lr_scheduler="${LR_SCHEDULER}" \

    --max_train_epochs $epocs \

    --sdpa \

    --mixed_precision=$mixed_precision \

    --save_precision=$save_precision \

    --save_every_n_steps=$save_every_n_steps \

    --save_model_as="safetensors" \

    --seed=1337 \

    --flip_aug \

    --enable_bucket \

    --no_metadata \

    --max_bucket_reso=$max_resolution_bucket \

    --min_bucket_reso=256 \

    --shuffle_caption \

    --logging_dir=logs \

    --log_prefix="${short_name}_lr${learning_rate}_" \

    --output_name="${FULL_NAME}" \

    --network_dim=$network_dim \

    --network_alpha=$alpha \

    --caption_extension ".txt" \

    --caption_tag_dropout_rate $tag_dropout \

    --caption_dropout_rate $caption_dropout \

    --network_dropout $dropout \

    --optimizer_type "AdamWScheduleFree" \

    --network_module=networks.lora_anima \

    --network_args "conv_dim=${conv_dim}" "conv_alpha=${conv_alpha}" \

    --bucket_reso_steps $bucket_reso_steps \

    --lr_warmup_steps $WARMUP_STEPS \

    --sample_every_n_steps $sample_every_n_steps \

    --sample_sampler "euler" \

    --sample_prompts "$sample_prompts_path" \

    --gradient_checkpointing \

    --max_token_length=$tokens \

    --scale_weight_norms 5 \

    --vae_batch_size 8 \

	--network_train_unet_only \

    --qwen3 "../../models/Qwen3-0.6B/model.safetensors" \

    --vae "../../models/Anima/split_files/vae/qwen_image_vae.safetensors" \

    --timestep_sampling="sigmoid" \

    --discrete_flow_shift=1.0

Custom training configs:

I have implemented a number of things into Kohys's training code that have been suggested to improve training, and kept the things that seemed to make improvements.

• drop out 75% of tags 5% of the time to hopefully improve short tag length results

• --no_flip_when_cap_matches: Prevent flipping images when certain tags exists like "sequence, asymmetrical, before and after, text on*, written, speech bubble" etc... This should help with text, and characters with asymmetrical features.

• --important_tags: move important tags to the beginning of the list, and sort them separately from the unimportant ones (suggested from NovelAI if I remember correctly).

• --tag_implication_dropout: Dropout similar tags to prevent the model from requiring them both to be present when generating. Like "breasts, big breasts" breasts will be dropped out 30-50% of the time. I used the tag implications csv from e621 as a base and added tags as needed. Even with 10%-15% tag dropout, some tag pairs were still being associated too often, this definitely made a difference. I think there were about 5k tags in total on the dropout list.

• 30% of the dataset is captioned with VLM's (Quen2, Qwen3, etc), as well as cleaning up many of the captions with custom scripts that correct common problems.

• Tags vs Captions: 70% of the time use tags, ~20% of the time use captions (if they exist), 10% of the time combine tags with captions in different orders.