SeeThrough Load LayerDiff Model

Every workflow in this pack starts here. SeeThrough Load LayerDiff Model is a loader node, and like most loader nodes it's boring in the best way: it fetches the SDXL-based layer-generation pipeline and hands it to the nodes downstream. But because the model it loads is the single biggest VRAM consumer in the entire See-through stack, the five toggles on this node are where you either fit the pack on your GPU or you don't. It's the one node in the pack worth actually reading the options on.

What it loads. The "LayerDiff 3D" model - layerdifforg/seethroughv0.0.2_layerdiff3d - a transparent-layer generation pipeline built on the same latent-transparency ideas as lllyasviel's LayerDiffusion, meaning it generates RGBA directly rather than matting afterward. It auto-downloads from HuggingFace on first use; several GB, so go make tea. The dropdown's second choice (24yearsold/..._nf4) is the 4-bit quantized variant for low-VRAM machines.

The inputs that matter:

• model - a dropdown that lists anything you've placed in ComfyUI/models/SeeThrough/ plus the HuggingFace defaults. You mostly leave this alone.
• quant_mode - none (default, bf16) or nf4 (4-bit, needs bitsandbytes). NF4 roughly halves the model's footprint - README's figure is ~8GB - at a quality cost you mostly don't see in the layers. If the bf16 model doesn't fit, this is a legitimately good trade.
• cache_tag_embeds - default true, leave it on. It pre-computes the tag embeddings once and then unloads the text encoders, saving ~2GB with zero speed impact. There is no reason to turn this off on any consumer GPU.
• group_offload - default false, and this is the "I'm desperate" switch. It moves model blocks on and off the GPU to cut peak VRAM (allocated VRAM drops to ~0.2GB), but it makes the run 2–3x slower, and it silently does nothing unless you have diffusers>=0.37.0 installed - check the console for the warning. Try resolution, quant_mode, and resolution_depth before you touch this.
• auto_download - default true. Set it to false if you're offline or want to force local-only loading; the node then errors instead of silently pulling from HuggingFace.
• vae_ckpt and unet_ckpt - optional paths to custom VAE/UNet checkpoints. These are for people experimenting with alternate weights, not for normal use. Leave them empty.

Output: layerdiff_model (type SEETHROUGH_LAYERDIFF_MODEL) goes into the layerdiff_model input of SeeThrough Generate Layers. That's the only thing it connects to.

Installing. This is part of the ComfyUI-See-through pack. Install via ComfyUI Manager (search "See-through") or:

cd ComfyUI/custom_nodes
git clone https://github.com/jtydhr88/ComfyUI-See-through.git
cd ComfyUI-See-through
pip install -r requirements.txt

Then restart ComfyUI. For offline setups, you can drop the model manually into ComfyUI/models/SeeThrough/ - the loader scans up to two levels deep for any folder containing model_index.json and lists it in the dropdown. Once it finds a local copy, it sets local_files_only=True and never phones home, even if the HF repo updates.

Where people get burned: the README promises "only 4 additional Python packages," but the shipped requirements.txt actually installs seven, including bitsandbytes, peft, and matplotlib - and group_offload specifically needs a diffusers version newer than the floor. So if you're getting a silent "skipping group_offload" warning, run pip install "diffusers>=0.37.0". And the first-run model download can time out on slow connections - rerunning the workflow resumes from the HF cache, so it's annoying rather than fatal.