ComfyUI
Node Graph
Every model call is a node with visible inputs and outputs. The workflow is a JSON file you can archive, share, and rerun unchanged.
Local and Private
The models run on your GPU. Evidence-derived images never leave the machine, which matters when the material is sensitive.
Model Agnostic
Text-to-image, image-to-image, depth estimation, and 3D body and object generation all run through the same graph, from official and community nodes.
Introduction
ComfyUI is a free, open-source, node-based application for local generative AI. You wire models together as a graph and run the graph. It is the workbench for the synthetic-asset side of visual investigation: generating illustrative references, estimating depth, and producing candidate 3D geometry from a single image. The method and its evidence rules are in the AI mesh generation technique, and the camera and perspective technique is what you verify any synthetic asset against.
The rule for this tool: the output is synthetic. It supports the work. It is not the evidence, and no dimension taken from it enters a finding.
Installation
- Download the application. The desktop app for Windows, macOS, and Linux is on the official site, and the source is on GitHub.
- Check your GPU. The models you can run are set by your VRAM. Record the GPU in your decision log.
- Install the node manager. ComfyUI-Manager adds the in-app installer for custom nodes. Then verify the version and record it. The current release is 0.36.0 (2026-09-15).
Usage
Choose the task· set the input and output
Decide what the run produces: a depth map, a synthetic image, or a candidate 3D mesh. What goes in: the input image (a working copy, never the original evidence file) or the text prompt. What comes out: the target artifact. Write both down before you build the graph.
Load the workflow· the JSON is the record
Load a workflow JSON (from the official learning series or your own library), or build the graph. Save the JSON and give it a case filename. The file is the reproducibility record: it stores every node, every setting, and every connection.
Install the models and nodes· name and hash
Install the model and the custom nodes the workflow needs. Record the model name and its hash in the decision log. A workflow without the exact model is not rerunnable.
Run with a fixed seed· deterministic output
Set the seed, the step count, and the guidance value, then run. A fixed seed makes the run deterministic: the same inputs, the same model, the same seed, the same output. If the result is wrong, change one parameter at a time and note the change.
Export and archive· artifact plus record
Export the artifact. What comes out: the output file, the workflow JSON, and the record line (model hash, seed, settings, date). Store all three together in the case file.
In an Investigation
Use the block below to document how you ran ComfyUI on the evidence.
Documentation: how to document ComfyUI
In Your Method Section
Guidance: Describe the purpose, the tool version, and the input data.
- Investigative Objective: To generate a synthetic artifact (depth map, reference image, candidate mesh) to support [specific analysis].
- Software Version: ComfyUI [Version] (e.g., 0.36.0), GPU [name, VRAM].
- Input Data: [Input image name or prompt], derived from [source document], Source Hash: [First 6 digits]. Input was a working copy.
In Your Decision Log
Guidance: Record the specific procedural parameters and integrity checks.
- Specific Settings: Workflow JSON filename, model name and hash, seed, steps, guidance value, any parameters changed from the base workflow.
- Assumptions: The generated result is a synthetic estimate. It is not a physical measurement and not direct evidence.
- Integrity Check: Re-ran the workflow once with the same seed to confirm the output is reproducible.
- Date Performed: [YYYY-MM-DD].
Verification & Mitigation
Guidance: How did you verify accuracy and mitigate bias?
- Verification: Compared the generated geometry or image against the visible cues in the input (perspective, scale relationships, known features).
- Bias Mitigation: Ran multiple seeds and checked the spread. Where the spread changed the conclusion, the artifact was discarded.
Example Methodology Statement
"A depth profile of the scene was estimated with ComfyUI 0.36.0 (GPU [X]) using the [model name, hash] model. The workflow [file.json] ran with seed [N], [steps] steps, guidance [G]. A repeat run with the same seed reproduced the output. The depth map is a synthetic estimate and was used for visualization only; no dimension from it entered the findings."
Into Blender
The artifacts from ComfyUI import into Blender as reference material.
- Meshes: import the generated mesh with the matching Blender importer (OBJ, glTF, or FBX, depending on the model). Place it in a collection clearly named as synthetic.
- Depth maps: use the depth map as a reference image for a plane, or as input to a meshing step in the AI mesh generation technique.
- Reference images: lay a generated image under a reconstruction as a perspective check, using the camera and perspective technique to test it.
The scale of a generated mesh is not metric. Do not measure against it.
Limitations
- Generative models invent detail that is not in the input. Hallucinated geometry looks as real as captured geometry.
- The scale and perspective of a generated model are relative, not metric.
- Results depend on the seed and the model version. The ecosystem moves fast, so pin the version, the model hash, and the workflow JSON, or the run is not rerunnable.
- The runs need a capable GPU, and large models exceed the VRAM of most consumer cards.
Related
Resources
- AI Mesh Generation, the technique and its evidence rules
- Camera and Perspective, for testing synthetic assets against the real viewpoint
Sources
- ComfyUI, official site with the desktop download
- ComfyUI documentation, official docs and interface guide
- ComfyUI releases on GitHub, current version 0.36.0 (2026-09-15)
- ComfyUI learning series, official video tutorials from basics to VFX