ComfyUI Node
ROCm VAE Decode
ROCM-optimized VAE Decode for AMD GPUs
ROCm VAE Decode
- samples
- vae
- IMAGE
◄tile_size768►
◄overlap96►
◄use_rocm_optimizationstrue►
◄precision_modeauto►
◄batch_optimizationtrue►
◄compatibility_modefalse►
◄enable_temporal_tilingauto►
◄temporal_chunk_size16►
◄temporal_overlap2►
◄last_frame_fixfalse►
CategoryROCm Ninodes/VAE
Inputs (12)
| Name | Type | Default | Description |
|---|---|---|---|
| samples | LATENT | The latent to be decoded. | |
| vae | VAE | The VAE model used for decoding the latent. | |
| tile_size | INT | 768256–2048 | Tile size. Larger values use more VRAM but are faster. |
| overlap | INT | 9632–512 | Overlap between tiles. Higher values reduce artifacts but use more VRAM. |
| use_rocm_optimizations | BOOLEAN | true | Enable ROCm-specific optimizations for AMD GPUs |
| precision_mode | COMBO | auto | Precision mode. 'auto' selects optimal for your GPU. |
| batch_optimization | BOOLEAN | true | Enable batch processing optimizations |
| compatibility_modeopt | BOOLEAN | false | Enable stock ComfyUI compatibility mode (disables all ROCm optimizations) |
| enable_temporal_tilingopt | COMBO | auto | For LTX/WAN videos: 'auto' enables temporal tiling for large outputs, 'enable' forces it on, 'disable' turns it off. |
| temporal_chunk_sizeopt | INT | 164–256 | Temporal tile size in LATENT frames. 16 latent = ~121 output frames for LTX. |
| temporal_overlapopt | INT | 21–8 | Overlap in LATENT frames between temporal tiles. Higher = smoother but more overhead. |
| last_frame_fixopt | BOOLEAN | false | Repeat last latent frame before decode, then discard extra output frames. Fixes end artifacts. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| IMAGE | IMAGE | — |