🔧 Flux LayerDiffuse Standalone Loader

Every transparent-Flux workflow in this pack starts here. FluxLayerDiffuseStandaloneLoader loads the TransparentVAE.pth checkpoint - the piece that actually knows how to emit an alpha channel - and hands it to the decoder as a TRANSPARENT_VAE. The "standalone" in the name is the point: it loads the transparent VAE directly from the checkpoint file instead of going through ComfyUI's normal VAE wrapper, which is the load path that trips up other LayerDiffuse setups.

The inputs

• vae (VAE, required) - your regular Flux VAE (ae.safetensors). The standalone transparent VAE builds on top of it, so both are needed.
• transparent_vae_checkpoint (STRING, default TransparentVAE.pth) - the filename. It looks for the file in models/vae/ first, then a few fallback locations including models/checkpoints/ and the pack's own models/ folder.
• alpha (FLOAT, default 300, range 0–1000) - the latent-transparency alpha offset that scales how strongly the alpha is baked into the latent space. This is the LayerDiffusion mechanism from the original paper: a latent-transparency offset folded into the model's latent manifold. Higher values push harder on the transparency signal. 300 is the pack's tuned default; if your output comes out too transparent (the "glass" look), this and the LoRA strength are the two knobs the README points at.
• latent_channels (INT, default 16, range 4–32) - Flux uses 16 latent channels, so leave it unless you know otherwise.
• dtype - bfloat16 (default), float16, or float32. bfloat16 is the sensible default on modern cards; float32 if you're chasing precision on a beefy setup and have VRAM to burn.

Output: transparent_vae (TRANSPARENT_VAE) → FluxLayerDiffuseDecoderSimple.

How it fits the workflow

The loader connects to the decoder, not to the sampler. The intended chain is:

FluxLayerDiffuseStandaloneLoader → FluxLayerDiffuseDecoderSimple → Save/Preview Transparent Image

and the KSampler's latents flow into the decoder separately. If you see an old workflow that passes a VAE straight into the decoder, that's the pre-standalone arrangement - the decoder still tolerates it but prints a warning telling you to route through this loader instead.

What you need before it works

The loader needs the actual files. TransparentVAE.pth ships bundled inside the pack's repo (README says so explicitly), so it's already sitting in the pack's models/ folder - but the loader's default name expects it in models/vae/, so either copy it there or point transparent_vae_checkpoint at wherever it lives. You also need the base Flux VAE (ae.safetensors) and, for the whole pipeline to mean anything, a Flux checkpoint, the layerlora.safetensors LoRA, and the Flux CLIP pair. If something's missing, the FluxLayerDiffuseInfo node's check_files action will tell you which one.

Install and troubleshooting

ComfyUI Manager → search "DenRakEiw Nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/DenRakEiw/DenRakEiw_Nodes
cd DenRakEiw_Nodes && pip install -r requirements.txt

then restart. The pack's requirements pull in diffusers 0.32.2, transformers, and peft, which the transparent VAE needs.

One honest word: this is a small author's port of a technique whose upstream tooling stopped being maintained in 2024–2025. When the pipeline misbehaves, check the file paths first (this is the most common loader failure - the checkpoint simply isn't where the node looks), then the CLIP type, then LoRA strength. The GitHub issues page is your support desk; there isn't a community behind it.