How to Build, Run, and Scale High-Quality Creator Workflows in ComfyUI

Creative and visualization teams today produce more assets, in more formats, with leaner teams. Generative AI can accelerate that work – compressing tasks that once took hours of manual effort into automated, repeatable pipelines.

ComfyUI is an open-source, node-based creative tool that runs locally on NVIDIA RTX GPUs. It connects image generation, video synthesis, and language models into pipelines that teams can customize and extend — without cloud dependencies or data leaving the client.

This guide walks through three production-ready workflows from the NVIDIA GenAI Creator Toolkit, adapted from NVIDIA’s GTC 2026 DLI course Create Generative AI Workflows for Design and Visualization in ComfyUI. Each workflow is standalone and runs locally on NVIDIA RTX.

What you’ll accomplish

By the end of this guide you will have:

1. Deconstructed an image into separate layers—foreground, midground, and background, each with a clean alpha mask, ready to bring into any compositing tool
2. Removed an unwanted object from a photo—and had AI fill in what should be there, seamlessly, without touching the rest of the image
3. Turned a photo into a textured 3D model (Windows only)—a 3D file in .GLB format you can open and use as a starting point for previs or layout.

Each workflow loads directly into ComfyUI as a pre-built node graph. You upload an image, click “Run,” and get results.

What to know before starting

• Basic comfort with a web browser and file uploads
• Beginner comfort with running scripts from the command line
• No ComfyUI experience required—the workflows are pre-built; you load them and run them
• Familiarity with image concepts (layers, masks, alpha channels) is helpful, but not required

Prerequisites

Hardware Requirements:

• Windows: NVIDIA RTX GPU with 24 GB of VRAM, with 32 GB or higher recommended
• Linux: NVIDIA RTX GPU with 32 GB of VRAM, with 48 GB or higher recommended
• 150 GB available disk space for the models these workflows require

Software Requirements:

• Required: Windows or Ubuntu Linux Operating System
• Required: Git (install here)
• Required: ComfyUI (Desktop App for Windows or manual install for Linux)
• Required for 3D Object workflow (module 08):  huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m, log in, and click “Agree and access repository”

Time and risk

• Estimated time: 30 minutes hands-on (plus model downloads of >50 GB on first run)
• Risk level: Low
  • All three workflows are read-only with respect to your input image—your original file is never modified.
  • Generated outputs are saved to ComfyUI’s output/ directory on your machine.
  • Rollback: The installer only adds files to your ComfyUI installation (custom nodes, workflows, models). To undo, delete the added folders from custom_nodes/ and the downloaded models from models/.

Detailed instructions

Step 1. Install the workflows

Windows—Open a command prompt (search for “cmd” in the Start Menu) and run:

(Tip: Don’t forget to visit huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m, log in, and click “Agree and access repository” in advance.)

git clone https://github.com/NVIDIA/NVIDIA-GenAI-Creator-Toolkit
cd NVIDIA-GenAI-Creator-Toolkit
install.bat C:\path\to\ComfyUI --modules 02,03,08

Pass your ComfyUI installation location—the folder you chose during Desktop App setup. It contains your .venv, models, and custom_nodes\ folders.

Not sure where it is? Check Desktop App Settings > About > Arguments: –base-directory C:\path\to\your\installation-location

Linux – complete instructions here: NVIDIA-GenAI-Creator-Toolkit/LINUX_COMFYUI_INSTALLATION.md at main · NVIDIA/NVIDIA-GenAI-Creator-Toolkit:

git clone https://github.com/NVIDIA/NVIDIA-GenAI-Creator-Toolkit
cd NVIDIA-GenAI-Creator-Toolkit
bash install.sh /path/to/ComfyUI --modules 02,03

The installer takes care of everything: It downloads the AI models (~150 GB total), installs the required plugins, and copies the workflows into ComfyUI. It can take 30–60 minutes depending on your internet connection.

When it’s done, start ComfyUI.

Step 2. Decompose an image into layers

What you’ll make: A photo broken into separate PNG layers—foreground, midground, and background—each with clean transparency. Bring them straight into Photoshop, After Effects, DaVinci Resolve, or any tool that supports layers.

Open the workflow: In ComfyUI, click Browse Templates (top menu) and find 02-image-deconstruction. under the NVIDIA Creative Workflows section. Click it to open.

Run it:

1. In the Load Image box, upload a photo with a clear subject—a portrait, a product, an architectural shot, or a scene with things in the foreground and background.
2. Click Run.

That’s it. No masking needed — the AI reads the depth of the image automatically.

What to expect: 3–4 PNG files, one per layer, each with transparent areas where the other layers were. The background layer is filled in where the foreground used to be.

(Tip: Photos with clear depth work best—a person in front of a wall, a product on a table, a building against a sky. Flat, uniform images won’t separate into interesting layers because there’s no depth for the AI to read.)

Additional Details: NVIDIA-GenAI-Creator-Toolkit/workflows/02-image-deconstruction at main · NVIDIA/NVIDIA-GenAI-Creator-Toolkit

Going further

Pair this with a motion graphics workflow—animate each layer separately for a parallax effect, or adjust color grading per-layer to change mood without touching the whole image.

Step 3. Remove an object from a photo

What you’ll make: A clean version of any photo with an unwanted element—a person in the background, a logo, a piece of clutter—removed and replaced with what belongs there.

Open the workflow: In ComfyUI, click Browse Templates and find 03-targeted-inpainting. Click it to open.

Run it.

1. In the Load Image box, click to upload the photo you want to edit.
2. Right-click the image thumbnail → Open in Mask Editor. Use the brush to paint white over whatever you want removed. White means “change this.” Black means “leave this alone.” Click Save when done.
3. Click Run (or Queue Prompt).

What to expect: Your photo with the painted area cleanly removed and filled in. Results are saved to ComfyUI’s output/ folder.

Tip: Paint a little larger than the object itself—giving the AI some extra border usually produces a cleaner result. If the fill doesn’t look right, try adding a description of what should be there (like “wooden floor” or “plain white wall”) in the text prompt box.

If nothing seems to change: You may have forgotten to paint the mask. Go back, right-click the image, and paint before running again.

Additional Details: NVIDIA-GenAI-Creator-Toolkit/workflows/03-targeted-inpainting at main · NVIDIA/NVIDIA-GenAI-Creator-Toolkit

Going further

This workflow can also add things. Describe an object in the text prompt, paint where you want it to appear, and the AI will generate it into your scene—matching the lighting and style of the surrounding image. 

Step 4. Turn a photo into a 3D model

Windows only. This workflow requires Windows. Linux support is not currently available.

What you’ll make: A textured 3D model (GLB file) from a single photo of any object. Open it in Blender, a game engine, or any 3D app for previs, layout, or further work.

Before you run — one extra step: The AI model for this workflow requires a free account approval from Meta. Visit huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m, log in, and click Agree and access the repository. Approval typically takes 24–48 hours. The installer will remind you before it tries to download.

Open the workflow: In ComfyUI, click Browse Templates and find 08-image-to-3d.

Run it:

1. In the Load Image box, upload a photo of an object — a piece of furniture, a prop, a shoe, a sculpture. Clear subjects on plain backgrounds work best for a first try.
2. Click Run.

What to expect: A 3D model that matches the shape and surface of your photo. The workflow shows you a rotating preview directly in ComfyUI. The GLB file is also saved to output/ — open it in Blender with File > Import > glTF.

(Tip: The cleaner your photo, the better the model. Shoot against a plain background, make sure the whole object is visible, and avoid extreme angles. Front-facing or 3/4 angle shots give the AI the most to work with.)

Additional Details: NVIDIA-GenAI-Creator-Toolkit/workflows/08-image-to-3d at main · NVIDIA/NVIDIA-GenAI-Creator-Toolkit

Step 5. Cleanup

When you’re done, stop ComfyUI to free GPU memory and system resources.

Windows: Close the ComfyUI Desktop App from the system tray or taskbar. Linux: Press Ctrl+C in the terminal where ComfyUI is running.

Your generated outputs remain in ComfyUI’s output/ directory, and all downloaded models stay in ComfyUI/models/. Nothing is deleted.

To free up disk space, remove a module’s model files with --clean. Models shared with other installed modules are kept automatically.

# Windows:
install.bat C:\path\to\ComfyUI --clean --modules 02,03

# Linux:
bash install.sh /path/to/ComfyUI --clean --modules 02,03

To restart and pick up where you left off, launch ComfyUI again and reload the workflow.

What’s Next

These three workflows are the entry point to a larger collection of pipelines. Here’s what’s waiting when you’re ready to go deeper:

1. AI Prompt Enhancement: Build an agent that rewrites rough creative prompts into optimized, model-ready descriptions, then generates an image from the result so you can see the difference immediately.

2. 360° Environment and HDRI Generation: Take any photo and generate a production-ready HDRI light for Unreal, Blender, V-Ray, or Arnold. No on-set capture required.
   
3. Playblast to Cinematic Video: Feed a flat gray-box render from Blender and a single style reference image. Get back a fully-lit, stylized video shot — swap the style image to explore multiple art directions without re-animating.

5. PBR Material Pipeline: Extract clean, seamless textures from any image and decompose them into a full PBR material set: albedo, roughness, metallic, normal, and height maps.

To install any additional module, run the installer again with the module number from here NVIDIA-GenAI-Creator-Toolkit/workflows—already installed nodes are skipped automatically:

Windows:

install.bat C:\path\to\ComfyUI --modules 01,04,05

Linux:

bash install.sh /path/to/ComfyUI --modules 01,04,05

The full DLI course covers all workflows with in-depth documentation, sample assets, and a capstone project that chains them together into a complete production pipeline.

Troubleshooting