SeeThrough Load Depth Model

SeeThrough Load Depth Model is the little brother of the two loader nodes in this pack. Where its sibling loads the multi-gigabyte LayerDiff model that generates the layers, this one loads the depth estimator - a fine-tuned Marigold - that works out what's in front of what. It's a smaller model, a smaller download, and a smaller VRAM footprint, and its options are the same shape as the LayerDiff loader's with a couple of knobs missing. Once it's loaded, you basically never think about it again.

What it loads. The dropdown defaults to layerdifforg/seethroughv0.0.1_marigold, a Marigold depth pipeline fine-tuned on anime. Marigold's whole trick (from the ETH Zurich group) is that it's generative - it reframes depth estimation as a denoising diffusion process on top of a Stable Diffusion prior, which is precisely why it copes with flat anime art that makes conventional photo-trained depth models throw up their hands. The second dropdown entry is the nf4 4-bit quantized variant. Auto-downloads from HuggingFace on first use - a much smaller pull than the LayerDiff model, but still a pull.

The inputs that matter:

• model - dropdown of local models in ComfyUI/models/SeeThrough/ plus the HF defaults. Leave it alone unless you've manually placed files.
• quant_mode - none (bf16, default) or nf4 (needs bitsandbytes). NF4 shrinks the depth model substantially; depth estimates are forgiving of quantization, so this is a low-pain way to free VRAM if your card is tight.
• cache_tag_embeds - default true. Here it caches the single empty text embedding the depth pipeline needs and unloads the text encoder to save VRAM. Leave it on.
• group_offload - default false. Same deal as its big sibling: dramatic VRAM savings at a large speed cost, and it requires diffusers>=0.37.0 or it silently skips (watch the console warning). On the depth model this matters less, since Marigold's peak is small next to the layer generation anyway.
• auto_download - default true. Flip to false for offline/local-only and the node errors instead of downloading.

Output: depth_model (type SEETHROUGH_DEPTH_MODEL), which plugs into the depth_model input of SeeThrough Generate Depth. That's its only job.

Installing. It ships in the ComfyUI-See-through pack - install the pack through 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

Restart ComfyUI and the node appears under the SeeThrough category. Offline users can place the model under ComfyUI/models/SeeThrough/ (any folder containing model_index.json, up to two levels deep, gets listed in the dropdown); once a local copy is found the loader goes local_files_only=True and never hits HuggingFace again.

Where people get burned: honestly, the depth loader is the least likely node in the pack to cause you grief - the failure modes live in the layer loader (VRAM) and the postprocess node (splitting). The one recurring gotcha is importing the node and expecting a model to appear instantly; the download happens lazily on the first run of the workflow, not at load time, so a slow connection can look like a hang. Let it sit. If you're on a tight GPU and skipped reading the sibling loader's group_offload warning, note that the resolution_depth setting on SeeThrough Generate Depth is usually a better first lever than quantizing this model at all.