<p>Updates will be mirrored on both Hugging Face and Civitai.</p><h2>Introduction</h2><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/23476/the-idolmster-cinderella-girls-starlight-stage-style-90-characters"><strong>It has been shown that multiple characters can be trained into the same model</strong></a>. A harder task is to create a model that can generate multiple characters simultaneously without modifying the generation pipeline. This document describes a simple technique that has been shown to help generating multiple characters in the same image.</p><h2>Method</h2><pre><code>Requirement: Sets of single-character images
Steps:
1. Train a multi-concept model using the original dataset
2. Create an augmentation dataset of joined image pairs (composites) from the original dataset
3. Train on the augmentation dataset
</code></pre><h2>Experiment</h2><h3>Setup</h3><p>3 characters from the game Cinderella Girls are chosen for the experiment. The base model is <code>anime-final-pruned</code>. It has been checked that the base model has minimal knowledge of the trained characters. Images with multiple characters were removed (edit: after training it was found that some unrelated background characters were not removed).</p><p>For the captions of the joined images, the template format <code>CharLeft/CharRight/COMPOSITE, TagsLeft, TagsRight</code> is used.</p><p>A LoRA (Hadamard product) is trained using the config file below:</p><pre><code>[model_arguments]
v2 = false
v_parameterization = false
pretrained_model_name_or_path = "Animefull-final-pruned.ckpt"

[additional_network_arguments]
no_metadata = false
unet_lr = 0.0005
text_encoder_lr = 0.0005
network_module = "lycoris.kohya"
network_dim = 8
network_alpha = 1
network_args = [ "conv_dim=0", "conv_alpha=16", "algo=loha",]
network_train_unet_only = false
network_train_text_encoder_only = false

[optimizer_arguments]
optimizer_type = "AdamW8bit"
learning_rate = 0.0005
max_grad_norm = 1.0
lr_scheduler = "cosine"
lr_warmup_steps = 0

[dataset_arguments]
debug_dataset = false
# keep token 1
# resolution 640*640

[training_arguments]
output_name = "cg3comp"
save_precision = "fp16"
save_every_n_epochs = 1
train_batch_size = 4
max_token_length = 225
mem_eff_attn = false
xformers = true
max_train_epochs = 40
max_data_loader_n_workers = 8
persistent_data_loader_workers = true
gradient_checkpointing = false
gradient_accumulation_steps = 1
mixed_precision = "fp16"
clip_skip = 2
lowram = true

[sample_prompt_arguments]
sample_every_n_epochs = 1
sample_sampler = "k_euler_a"

[saving_arguments]
save_model_as = "safetensors"
</code></pre><p>For the second stage of training, the batch size was reduced to 2 and the resolution was set to 768 * 768 while keeping other settings identical. The training took less than 2 hours on a T4 GPU.</p><h3>Results</h3><p>(see preview images)</p><h2>Limitations</h2><ul><li><p>This technique doubles the memory/compute requirement</p></li><li><p>Composites can still be generated despite negative prompting</p></li><li><p>Cloned characters seem to become the primary failure mode in place of blended characters</p></li><li><p>Can generate 2 characters but not 3 for some reason</p></li></ul><h2>Related Works</h2><p>Models been trained on datasets based on anime shows have <a target="_blank" rel="ugc" href="https://civarchive.com/models/21305/"><strong>demonstrated</strong></a> multi-subject capability. Simply using concepts distant enough such as <code>1girl, 1boy</code> <a target="_blank" rel="ugc" href="https://civarchive.com/models/17640/"><strong>has also been shown to be effective</strong></a>.</p><h2>Future work</h2><p>Below is a list of ideas yet to be explored</p><ul><li><p>Synthetic datasets</p></li><li><p>Unequal aspect ratios</p></li><li><p>Regularization</p></li><li><p>Joint training instaed of sequential</p></li></ul>

Guide: 1to2: Training Multiple-Subject Models using only Single-Subject Data (Experimental)

<p>Result before second stage training</p>

Files with SHA256: aafcdbfc6da661527df830a65e9e766f91a83a80a5526acf5eb2832c255bb453