v1.5-hf/ffSplit two concepts into separate files, consequencially increasing amout of pictures in both datasets. - Feet first is pretty much the same as in 1.0, but now with "peaking out" and "head only" variety (pretty much like with Pony model, that some may have seen on this site). - Head first is unstable. Separating and isolating it seem to not have enough of an effect. May say that this is expected, because model is "not used to" people without upper body in frame. And all generative models are pretty bad at depiction of implied/obscured/not visible things, not only Anima. v1.0 First try. Dataset has 44 images of questionable quality. All tagged by hand."head-firs/feet-first" seem to be ignored. As I understand human figure is too strong of a shape and that was the reason why simillar loras for illustrious are split. Stage "half-body/quarter-body" is somewhat working, but weak. More of a suggestion rather than clear trigger. (update): apparently, head first works better without tag at all. My only guess is that word "head" in any form is already too strong in the model itself. Tag "unbirthing ambiguous" was ment to be for cases when there is no visible hints for gender, but turned to be too wague. Probably will generate some kind of amorphous blob, but I'm not an expert in prompting Anima models. Angles of view are lacking, most of dataset images were from the front. For everyone, who wants better future versions - post your good pics with this lora in resources, coment links for good looking references from other resources and share your experience on this theme. Trained with citron trainer. Params are as follows:<pre><code>{</code></pre><pre><code>network_dim = 24
network_alpha = 24
network_train_unet_only = true
learning_rate = 0.0001
optimizer_type = "AdamW8bit"
optimizer_args = [ "weight_decay=0.1", "betas=[0.9, 0.99]",]
lr_scheduler = "cosine_with_restarts"
lr_scheduler_num_cycles = 1
lr_warmup_steps = 100
max_train_epochs = 10
train_batch_size = 1
gradient_accumulation_steps = 1
max_grad_norm = 1.0
seed = 42
timestep_sampling = "sigmoid"
discrete_flow_shift = 1.0
qwen3_max_token_length = 512
t5_max_token_length = 512
mixed_precision = "bf16"
gradient_checkpointing = true
cache_latents = true
cache_text_encoder_outputs = true
vae_chunk_size = 64
vae_disable_cache = true
output_name = "unbirth_anima_v1"
save_model_as = "safetensors"
save_precision = "bf16"
save_every_n_epochs = 1
save_last_n_epochs = 4
shuffle_caption = false
caption_extension = ".txt"
noise_offset = 0.03
multires_noise_discount = 0.3
}
</code></pre> Base One can say that this was done by comission, but no one payed me anything, sadly. Lora by itself is quite self explanatory. Main purpose is to depict a stage in process of unbirthing - absorbing/sucking/consuming of one character be another's vagina and uterus.

Un-birth [Anima]

Files with SHA256: bf47580ae6276865eead470aab00a9b60288f57c8dd5b34b8c38b1ac8cc7f8cd