Minimax H3 test LoRA for pussy spread
This is just a test LoRA, trained on only 10 pussy spread clips, so please don't expect great results.
Training setting:
job: "extension"
config:
name: "minimax_v3"
process:
- type: "diffusion_trainer"
training_folder: "/path/to/output"
sqlite_db_path: "./aitk_db.db"
device: "cuda"
trigger_word: null
performance_log_every: 10
network:
type: "lora"
linear: 16
linear_alpha: 16
conv: 16
conv_alpha: 16
lokr_full_rank: true
lokr_factor: -1
network_kwargs:
ignore_if_contains:
- "adaln_proj"
save:
dtype: "bf16"
save_every: 250
max_step_saves_to_keep: 4
save_format: "diffusers"
push_to_hub: false
datasets:
- folder_path: "/path/to/datasets/my_dataset"
mask_path: null
mask_min_value: 0.1
default_caption: ""
caption_ext: "txt"
caption_dropout_rate: 0
cache_latents_to_disk: true
is_reg: false
network_weight: 1
resolution:
- 1024
- 768
controls: []
shrink_video_to_frames: true
num_frames: 39
flip_x: false
flip_y: false
num_repeats: 1
do_i2v: true
fps: 24
auto_frame_count: true
train:
batch_size: 1
bypass_guidance_embedding: false
steps: 3000
gradient_accumulation: 1
train_unet: true
train_text_encoder: false
gradient_checkpointing: true
noise_scheduler: "flowmatch"
optimizer: "automagic3"
timestep_type: "shift"
content_or_style: "balanced"
optimizer_params:
weight_decay: 0.0001
unload_text_encoder: false
cache_text_embeddings: true
lr: 0.0001
ema_config:
use_ema: false
ema_decay: 0.99
skip_first_sample: false
force_first_sample: false
disable_sampling: false
dtype: "bf16"
diff_output_preservation: false
diff_output_preservation_multiplier: 1
diff_output_preservation_class: "person"
switch_boundary_every: 1
loss_type: "mse"
do_guidance_loss: true
guidance_loss_target: 3.5
audio_loss_multiplier: 1
do_differential_guidance: true
differential_guidance_scale: 4
logging:
log_every: 1
use_ui_logger: true
model:
name_or_path: "Comfy-Org/MiniMax-H3"
quantize: true
qtype: "convrot8"
quantize_te: true
qtype_te: "nvfp4"
arch: "minimax_h3"
low_vram: false
model_kwargs: {}
compile: false
layer_offloading: false
layer_offloading_text_encoder_percent: 1
layer_offloading_transformer_percent: 1
assistant_lora_path: "ostris/minimax_h3_training_adapter/minimax_h3_training_adapter_v1.safetensors"Description
Test LoRA trained on only 10 videos
FAQ
Comments (10)
very epic! Looks like her uretha gets lost in a lot of spreading loras.
>This is just a test LoRA, trained on only 10 pussy spread clips
Is undertrained
@anonlinuxmint962 2,000 steps should be enough. It's more like the training set lacks diversity. The LoRA will fail if pussy isn't large enough in the frame.
Does this only do hairy pussies?
This is i2v lora, so depends on the input image. The 10 clips in the training set are all hairless.
@salttaro Its impressive you pulled it off with just 10 clips! Did you actually train it in just I2V mode? I'm sorry for asking - feel free to ignore but, would you mind sharing your training settings? Also it would be cool to have a lora that actually teaches all pussy physics. Not just what it looks like on its own at different angles nor what it looks like when being spread - but also how it reacts to touch/object insertion/object removal.
@Winchester99 I've posted the training settings in the description of this LoRA. Your idea is pretty good. Collecting video clips as a training set showing how pussy reacts to touch, object insertion, or object removal might be a tough process.
@salttaro thanks for all your effort, but i feel if we help each other out, finding the touching, insertion and stretching, would be many people and many clips ready. My friend has JAV collection, he would easily find 10 clips of each haha...
This is exactly what we need👍