ComfyUI-See-through (Custom Fork)

A fork of ComfyUI-See-through by @jtydhr88, adding a custom node with the option to skip the head detail inference stage for faster processing.

中文說明

What's New in This Fork

v1.2.8 — Issue #5

• New node SeeThrough Load Source: same dropdown as ComfyUI's LoadImage, plus outputs source_filename for preserving the original filename in PSD output.
• SeeThrough Save PSD now accepts optional original_image + source_filename inputs, automatically includes the original input image as a visible base layer in the generated PSD, and uses the source filename in output filenames when available.
• PSD layer structure is now grouped: Original (visible, bottom), Parts (hidden), Runs (hidden, grouped-PSD mode only) — opens the PSD on the original so you can toggle groups to edit specific parts.

v1.2.4 — Bug Fixes

• Fix: Depth model fails to download — HuggingFace repo ID updated — The default Marigold depth model repo 24yearsold/seethroughv0.0.1_marigold was moved to layerdifforg/seethroughv0.0.1_marigold. Updated the default repo ID so depth model auto-download works again. (#3)

v1.2.3 — Bug Fixes

• Fix: PSD download fails with "Failed to load ag-psd bundle from any path" — ComfyUI's new frontend loads extensions via ES module import(), causing document.currentScript to be null. Replaced with import.meta.url for reliable bundle path resolution regardless of install folder name. (#1)
• Fix: Nodes fail to load in fresh environments — The upstream Marigold module imports matplotlib at module level for an optional visualization function, causing ModuleNotFoundError on environments without matplotlib. Changed to lazy import so nodes load without requiring matplotlib.

v1.2 — Spine Export, Auto-Fill & All Runs PSD

• Spine Export Nodes — New Layer Rename, Layer Filter, and Export Spine nodes for Spine 2D animation preparation.
• Auto-Fill Missing Layers — Enable auto_fill on GenerateLayers (Custom) to automatically re-run inference up to 5 times, filling missing layers and upgrading low-quality layers by comparing against the original image.
• All Runs PSD — When auto_fill is enabled, a new "Download All Runs PSD" button appears on Save PSD. It creates a single PSD with group folders for each run, so you can manually compare and pick layers.
• PSD Download Buttons — Save PSD now has 3 buttons:
  • Download PSD (green) — best layers after auto-fill selection
  • Download Depth PSD (purple) — depth maps
  • Download All Runs PSD (orange) — all runs grouped by folder (requires auto_fill)

SeeThrough Generate Layers (Custom)

A new node SeeThrough_GenerateLayers_Custom that adds one parameter compared to the original SeeThrough Generate Layers:

Parameter	Default	Description
enable_head_detail	true	v3 models only: toggle the head detail inference stage on/off
auto_fill	false	Auto-fill missing layers: re-runs inference (up to 5 times) until all expected layers are generated (v3+head=24, v3 body=13, v2=19)
min_alpha_coverage	0.01	Minimum alpha coverage ratio to consider a layer valid. Only used when auto_fill is enabled

How It Works

The v3 See-through model runs in two inference stages:

1. Body stage — Generates 13 body-level layers (front hair, back hair, head, neck, neckwear, topwear, handwear, bottomwear, legwear, footwear, tail, wings, objects)
2. Head stage — Crops the head region from stage 1, upscales it, and runs a second inference pass to generate 11 fine-grained head layers (headwear, face, irides, eyebrow, eyewhite, eyelash, eyewear, ears, earwear, nose, mouth)

Each stage is a full diffusion pipeline call. By setting enable_head_detail = false, the entire head stage is skipped (no GPU computation), saving approximately 50% of the total inference time.

This is useful when you only need body-level decomposition and don't require fine-grained facial features.

Multi-Run Auto-Fill

The diffusion model is stochastic — each run may produce slightly different results. Sometimes a layer (e.g., face or hand) is missing or nearly empty in one run but present in another.

Enable auto_fill to automatically re-run inference until all expected layers are generated with good quality:

1. Run 1 uses the original seed — this is the primary result
2. Each layer is compared against the original image to compute a similarity score (0~1)
3. Layers that are missing (alpha coverage below threshold) or have low similarity (< 0.85) trigger additional runs
4. Run 2 uses seed + 1, Run 3 uses seed + 2, etc.
5. For each layer, the version with the highest similarity to the original is kept
6. The process repeats up to 5 runs or until all layers have good similarity

This means even if Run 1 generates a face layer, if Run 2 produces a face that better matches the original image, Run 2's version will be used automatically.

Models are loaded to GPU only once across all runs — the overhead is only the additional diffusion time, not model loading.

Note: For v2 models, this toggle has no effect since v2 uses a single-stage inference.

v1.1 — Synced with Upstream

This fork has been synced with upstream v0.2.2, incorporating the following improvements:

• VRAM Offloading — Models now stay on CPU and are moved to GPU only during inference, then offloaded back. This significantly reduces VRAM usage, making it possible to run on GPUs with 8GB or less.
• Text Encoder Offloading — Text encoders are loaded to GPU for prompt encoding, then offloaded before loading UNet+VAE, so they never compete for VRAM at the same time.
• Marigold Compatibility Fix — Fixed resize call for torchvision >= 0.23 which introduced stricter InterpolationMode checks.
• Web Fixes — Subpath deployment support; ag-psd bundle 404 fix on case-sensitive filesystems.
• Custom Node Optimization — When enable_head_detail = false, head text encoding is also skipped (not just diffusion), further reducing GPU memory usage.

All Nodes

Node	Description
SeeThrough Load LayerDiff Model	Load the LayerDiff SDXL pipeline
SeeThrough Load Depth Model	Load the Marigold depth estimation pipeline
SeeThrough Generate Layers	Original layer generation (all stages, all layers)
SeeThrough Generate Layers (Custom)	Layer generation with enable_head_detail toggle
SeeThrough Generate Depth	Depth map estimation per layer
SeeThrough Post Process	Left/right splitting, hair clustering, color restoration
SeeThrough Save PSD	Export layers as PNGs + metadata; download Best PSD, Depth PSD, or All Runs PSD via browser
SeeThrough Layer Rename	Rename layer tags to Spine-friendly names (customizable)
SeeThrough Layer Filter	Include/exclude specific layers before export
SeeThrough Export Spine	Export layers as a Spine 2D skeleton project (JSON + images)

Spine Export Workflow

For Spine animation preparation, connect:

PostProcess → Layer Rename (optional) → Layer Filter (optional) → Export Spine

Layer Rename

Maps internal tags to Spine-friendly names. Has built-in defaults for all tags. The custom_mapping_json field is optional — leave it empty to use defaults.

When to use it:

• You want clean, readable names in Spine (e.g. front-hair instead of hairf)
• Your team has a naming convention and you need custom names

Built-in default mapping (partial list):

Original tag	→ Renamed to
hairf	front-hair
hairb	back-hair
eyel	eye-left
eyer	eye-right
handwearl	handwear-left
handwearr	handwear-right
earl	ear-left
earr	ear-right
topwear	topwear (unchanged)
face	face (unchanged)

Tags that already have clean names (e.g. face, head, nose) are kept as-is.

Custom mapping example: To override specific names, enter a JSON object in custom_mapping_json:

{
  "hairf": "bangs",
  "hairb": "back-hair",
  "topwear": "shirt",
  "bottomwear": "skirt",
  "handwearl": "left-glove",
  "handwearr": "right-glove"
}

Only the tags you specify in the JSON will be overridden — all other tags still use the built-in defaults. Invalid JSON is ignored with a warning.

Layer Filter

Removes unwanted layers using include or exclude mode. All available tags are pre-filled by default — delete the ones you don't need. Enter one tag per line.

Tip: If Layer Rename is connected before Layer Filter, use the renamed tag names (e.g. front-hair). If not using Layer Rename, use original tags (e.g. hairf).

Export Spine

Outputs a folder with a configurable output path (defaults to ComfyUI output directory):

• {prefix}.json — Spine skeleton file (open directly in Spine editor)
• images/ — cropped PNG files for each layer
• Set output_path to export to a custom directory (e.g. D:/my_project/spine_assets)

Coordinates are automatically converted from image space (Y-down) to Spine space (Y-up, origin at bottom-center). Draw order follows depth ordering from PostProcess.

PSD Import vs JSON Export — Which Should I Use?

Spine Professional (3.6+) can import PSD files directly, so you may wonder whether this JSON export is needed. Here's the comparison:

	Save PSD → Spine PSD Import	Export Spine (JSON + images)
Spine version	Professional 3.6+ only	All versions (Essential + Professional)
Layer positioning	Automatic	Automatic (coords pre-converted)
Layer naming	Depends on PSD layer names	Controllable via LayerRename
Layer filtering	Must hide/delete in PSD first	Built-in LayerFilter node
Iteration	Re-import PSD to update images	Re-export to update
Bone hierarchy	Not auto-created	Not auto-created
Best for	Spine Professional users who want a quick start	Spine Essential users, or teams wanting pre-filtered/renamed layers in an automated pipeline

Recommendation:

• Spine Professional users → Use Save PSD and import via Spine's built-in PSD import. It's the simplest workflow.
• Spine Essential users → Use Export Spine, as Essential does not support PSD import.
• Automated pipelines → Use Export Spine with LayerRename + LayerFilter for consistent, pre-processed output.

Available layer tags (after LayerRename, 38 tags)

Category	Tags
Hair	front-hair, back-hair
Head	head, headwear
Face	face, nose, mouth
Eyes	eye-left, eye-right, eyewear
Eye detail	irides, irides-left, irides-right, eyebrow, eyebrow-left, eyebrow-right, eye-white, eye-white-left, eye-white-right, eyelash, eyelash-left, eyelash-right
Ears	ears, ear-left, ear-right, earwear
Body	neck, neckwear, topwear, bottomwear
Limbs	handwear, handwear-left, handwear-right, legwear, footwear
Other	tail, wings, objects

If not using LayerRename, use original tags: hairf, hairb, eyel, eyer, handwearl, handwearr, earl, earr, etc.

Installation

Clone this repository into your ComfyUI custom_nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/tackcrypto1031/tk_seethrough.git

Install dependencies:

cd tk_seethrough
pip install -r requirements.txt

Restart ComfyUI. The nodes will appear under the SeeThrough category.

Models

Models are downloaded automatically from HuggingFace on first use:

Model	HuggingFace Repo	Purpose
LayerDiff 3D	layerdifforg/seethroughv0.0.2_layerdiff3d	SDXL-based transparent layer generation
Marigold Depth	layerdifforg/seethroughv0.0.1_marigold	Fine-tuned monocular depth for anime

You can also download models manually and place them in ComfyUI/models/SeeThrough/.

Usage

1. Add SeeThrough Load LayerDiff Model and SeeThrough Load Depth Model
2. Add SeeThrough Generate Layers (Custom) — connect both models and a Load Image node
3. Uncheck enable_head_detail if you want faster processing without head detail layers
4. Connect to SeeThrough Generate Depth → SeeThrough Post Process → SeeThrough Save PSD
5. Run the workflow and click Download PSD to export

For Spine export: Replace step 4's Save PSD with Layer Rename → Layer Filter → Export Spine. Open the output JSON in Spine editor.

Acknowledgements

This project is a fork of ComfyUI-See-through by @jtydhr88. Huge thanks for creating the original ComfyUI integration.

The underlying research is See-through by shitagaki-lab. Paper: arxiv:2602.03749 (Conditionally accepted to ACM SIGGRAPH 2026)

PSD generation uses ag-psd in the browser.

License

MIT