🛑Work in progress🛑

(Alpha release) I'm not sure this will be interesting to anyone.

• WORKFLOW: https://civarchive.com/models/2516563/wan-with-ltxv-23-audio

• Not designed for oral sex

  • I tried nothing more confusing or disturbing than hearing "gawk gawk" or gagging in an anal video.

  • Check out my deepthroat lora it may work for adding audio, confirmed to work.

  • If a 1GB lora is to much I may spend sometime to create a lightweight BJ audio lora.

Create sex audio for previously created videos or in addition to LoRAs that lack audio. Three main additions to the base model: clapping cheeks, improved moaning/heavy breathing, and wetness sounds.

This is a purely experimental LoRa addressing a common gap in many videos. It uses video-to-audio cross-attention to generate audio, meaning text prompts aren't critical but can still provide influence.

Tags used

- skin slapping against skin 
- clapping cheeks
- wet vagina
- The woman moans
- The woman is breathing heavy

Extra Information

I've tested with dev and distill the best results are from Dev.

• Best Samplers I've found - res_2s, er_sde

• Audio will sync to visual movement naturally

LoRa Creator info

Stand out info

• Rank 16 (might be a little to small)

• --lora_target_preset full for cross-attention

• -ltx2_mode av

• Separate audio learn rate

accelerate launch --num_cpu_threads_per_process 8 --mixed_precision bf16 \
  ltx2_train_network.py --sdpa \
  --ltx2_checkpoint /ai/comfyui/models/checkpoints/ltx-2.3-22b-dev.safetensors \
  --dataset_config ~/datasets/sex-audio/ltx_dataset_config.toml \
  --mixed_precision bf16 \
  --optimizer_type adamw8bit \
  --learning_rate 5e-5 \
  --gradient_checkpointing \
  --max_data_loader_n_workers 8 \
  --persistent_data_loader_workers \
  --network_module networks.lora_ltx2 \
  --network_dim 16 --network_alpha 16 \
  --timestep_sampling shifted_logit_normal \
  --discrete_flow_shift 1.0 \
  --max_train_steps 5000 --lr_scheduler constant --audio_lr 2.5e-5 \
  --max_grad_norm 1.0 \
  --save_every_n_steps 250 \
  --seed 42 \
  --logging_dir /ai/datasets/sex-audio/logs \
  --output_dir /ai/comfyui/models/loras/LTX2.3/sex-audio \
  --output_name sex-audio \
  --ltx2_first_frame_conditioning_p 1.0 \
  --caption_dropout_rate 0.1 --lora_target_preset full --ltx2_mode av