<h1 id="work-in-progress"><span style="color:rgb(250, 82, 82)">🛑Work in progress🛑</span></h1><h3 id="(alpha-release)-i'm-not-sure-this-will-be-interesting-to-anyone."><span style="color:rgb(253, 126, 20)">(Alpha release)</span> I'm not sure this will be interesting to anyone.</h3><ul><li><p><span style="color:rgb(250, 82, 82)">WORKFLOW: </span><a target="_blank" rel="ugc" href="https://civarchive.com/models/2516563/wan-with-ltxv-23-audio">https://civarchive.com/models/2516563/wan-with-ltxv-23-audio</a></p></li><li><p><span style="color:rgb(250, 82, 82)">Not designed for oral sex</span></p><ul><li><p>I tried nothing more confusing or disturbing than hearing "gawk gawk" or gagging in an anal video.</p></li><li><p>Check out my deepthroat lora it may work for adding audio, <span style="color:rgb(250, 82, 82)">confirmed to work</span>.</p></li><li><p>If a 1GB lora is to much I may spend sometime to create a lightweight BJ audio lora.</p></li></ul></li></ul><p></p><p>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.</p><p></p><p>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.</p><h2 id="tags-used">Tags used</h2><pre><code>- skin slapping against skin 
- clapping cheeks
- wet vagina
- The woman moans
- The woman is breathing heavy</code></pre><h2 id="extra-information">Extra Information</h2><p>I've tested with dev and distill the best results are from Dev.</p><ul><li><p>Best Samplers I've found - res_2s, er_sde</p></li><li><p>Audio will sync to visual movement naturally</p></li></ul><p></p><h2 id="lora-creator-info">LoRa Creator info</h2><p>Stand out info</p><ul><li><p>Rank 16 (might be a little to small)</p></li><li><p><code>--lora_target_preset full</code> for cross-attention</p></li><li><p><code>-ltx2_mode av</code></p></li><li><p>Separate audio learn rate</p></li></ul><pre><code>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
</code></pre><p></p>

Files with SHA256: a13beb8e021536198f17fb0b834a129d6c3f61248230ed9f29736a32068f23e4