VAE Lyra nodes for ComfyUI - Multi-modal VAE with Cantor fusion for SDXL text encoders.
cd ComfyUI/custom_nodes
git clone https://github.com/AbstractPhil/ComfyUI-Lyra.git
cd ComfyUI-Lyra
pip install -r requirements.txt| Model | Version | Fusion | CLIP Weights |
|---|---|---|---|
AbstractPhil/vae-lyra |
v1 | cantor | standard |
AbstractPhil/vae-lyra-sdxl-t5xl |
v1 | cantor | standard |
AbstractPhil/vae-lyra-xl-adaptive-cantor |
v2 | adaptive_cantor | standard |
AbstractPhil/vae-lyra-xl-adaptive-cantor-illustrious |
v2 | adaptive_cantor | illustrious |
Load model from HuggingFace Hub or local path.
Encode CLIP-L, CLIP-G, T5-XL embeddings to latent space.
Decode latent back to reconstructed embeddings.
Complete encode-decode cycle.
Display model configuration and learned parameters.
V1: Single encoder, single output projection. forward() returns 3 values.
V2 (adaptive_cantor): Per-modality encoders with hard masking between binding groups. forward() returns 4 values.
MIT