FoleyTune LoRAs
LoRA adapters trained with ComfyUI-HunyuanVideo-FoleyTune for video-to-audio generation on the HunyuanVideo audio transformer.
Adult content. Some adapters in this repository are trained on adult audio datasets. Browse subdirectories for details.
⚠️ Render CFG depends on the folder. Some adapters must be rendered at CFG = 1.0 (they saturate at higher guidance); others work at any normal CFG. Check the folder / its README before rendering.
Categories
| Folder | Description | Adapters |
|---|---|---|
type-a/ |
Adult audio LoRAs | 5 blowjob adapters |
CFG_1.0/ |
Render @ CFG 1.0 ONLY (saturate at higher CFG) | 1 sex/moaning + skin-slap merge |
standard_cfg/ |
Standard-CFG adapters (any normal guidance) | 1 blowjob |
The
CFG_1.0/andstandard_cfg/adapters use the newerall_blocks_sync_iotarget + Prodigy+ ScheduleFree (c=20) recipe (see each folder's README), and render best with TF32 off, raw output (no bandwidth extension). Thetype-a/config below is the older sweep.
Base Training Config
All adapters share this foundation unless noted in their category README:
- Base model: HunyuanVideo audio transformer
- LoRA target: all_attn_mlp (attention + MLP layers)
- Rank: 64, Alpha: 64
- Optimizer: Prodigy (adaptive learning rate)
- Schedule: Cosine with curriculum switching
- Precision: bf16
- Visual dropout: 0.5
Metrics
- PBC (Per-Band Correlation): frequency-band tracking accuracy. Higher is better.
- SC (Spectral Convergence): overall spectral fidelity. Lower is better.
- MCD (Mel-Cepstral Distortion): timbral accuracy. Lower is better.
- TV (Temporal Variance): dynamic range over time. Higher means richer variation.
Links
- GitHub: ComfyUI-HunyuanVideo-FoleyTune
Usage
Load in ComfyUI with the FoleyTune LoRA Loader node (supports both .safetensors and .pt), or manually:
import torch
from safetensors.torch import load_file
state_dict = load_file("type-a/typeA_blowjob_rank1_sigma07_15k_PBC0661_SC1058.safetensors")
# Apply to HunyuanVideo audio transformer with load_lora()