NVIDIA/NVIDIA-GenAI-Creator-Toolkit

NVIDIA/NVIDIA-GenAI-Creator-Toolkit

Description: A collection of generative ai workflows for creators.

Language: Python

License: NOASSERTION

Stars: 63

Forks: 12

Open issues: 1

Created: 2026-04-06T19:51:14Z

Pushed: 2026-05-04T17:45:44Z

Default branch: main

Fork: no

Archived: no

README:

ComfyUI Generative AI Workflows

Achieve professional creative control over 3D assets and motion for visualization, powered by modular generative AI pipelines on NVIDIA RTX.


Adapted from NVIDIA's GTC 2026 DLI course *Create Generative AI Workflows for Design and Visualization in ComfyUI* (DLIT81948). Each module is standalone — pick the pipelines that fit your work.

---

Requirements

• GPU: RTX 5090 (Windows)+ or RTX PRO 6000 (Linux) recommended to run all modules. Not all modules require as much VRAM, see the table below for requirements per module.
• Disk Space: 500GB for all modules. See table below for Disk Space requirements per module.
• OS: Windows 11 or Linux x86_64
• CUDA: 12.x (Windows: included with your NVIDIA driver; Linux: verify with nvidia-smi)
• Python: 3.10 or newer — 3.11 or 3.12 recommended (Windows Desktop App: bundled; Linux: python3.12-venv may be needed, see [LINUX_COMFYUI_INSTALLATION.md](LINUX_COMFYUI_INSTALLATION.md))
• HuggingFace account: Required for model downloads. The installer will prompt you to log in. Modules 07 and 08 also require accepting gated model agreements before downloading.
• Software: ComfyUI and Git

> New to ComfyUI? ComfyUI is a node-based generative AI interface — you connect model components visually to build pipelines. Each workflow in this repo is a pre-built pipeline you load and run.

---

Quick Start with Module 01

##### Linux Follow these step-by-step instructions: [LINUX_COMFYUI_INSTALLATION.md](LINUX_COMFYUI_INSTALLATION.md).

##### Windows Download and install the desktop app from https://www.comfy.org/download. Launch ComfyUI and complete its setup. Close ComfyUI when done. And continue below:

Clone the NVIDIA GenAI Creator Toolkit GitHub Repo

# Clone this repo
git clone https://github.com/NVIDIA/NVIDIA-GenAI-Creator-Toolkit
cd NVIDIA-GenAI-Creator-Toolkit

>Need Git on Windows? Download from git-scm.com/downloads and run the installer.

Install Module 01

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

install.bat C:\path\to\your\installation-location --modules 01

Open and Run Module 01's Workflow in ComfyUI

Start ComfyUI. From Windows open the Desktop app. Open the Templates window and scroll down to NVIDIA GenAI Creator Toolkit; open module 01.


Press the blue run button in ComfyUI and see your prompt improve!


> If ComfyUI shows a Missing Models dialog, the listed files need to be downloaded before generating. Re-run install.bat — already-downloaded models are skipped and only missing files are fetched.

---

The Modules in this Toolkit

Run the same script again with more module numbers.

# Windows:
install.bat C:\path\to\ComfyUI --modules 02,03
# Linux:
bash install.sh /path/to/ComfyUI --modules 02,03

| # | Workflow | Key Model(s) | Min. Rec. Windows / Linux VRAM | Disk Space | What It Does | |---|----------|------------|------|--------|-------------| | 01 | [LLM Prompt Enhancer](workflows/01-llm-prompt-enhancer/) | Gemma 3 via Ollama | 24 / 32 GB | ~65 GB | Build an AI agent that refines weak prompts into model-ready instructions | | 02 | [Image Deconstruction](workflows/02-image-deconstruction/) | Qwen Image Layered | 24 / 32 GB | ~51 GB | Split any image into foreground, midground, and background layers | | 03 | [Targeted Inpainting](workflows/03-targeted-inpainting/) | Qwen Image Edit 2511 | 24 / 32 GB | ~52 GB | Mask-and-patch editing — change only the pixels you select | | 04 | [Image → Equirectangular](workflows/04-image-to-equirectangular/) | Qwen Image Edit 2511 + Mickmumpitz QWEN-EDIT_360 LoRA | 24 / 32 GB | ~61 GB | Turn a single image into a seamless 360° panorama | | 05 | [Panorama → HDRI](workflows/05-panorama-to-hdri/) | Flux Dev Kontext + Exposure LoRAs | 24 / 32 GB | ~23 GB | Generate a production-ready HDRI from a panoramic image | | 06 | [Image Cut Out Time to Move](workflows/06-image-cut-out-time-to-move/) | Wan2.2 TTM + VideoPrep | 32 / 48 GB | ~77 GB | Trajectory-controlled video — define exactly when and where motion happens | | 07 | [Video to Video](workflows/07-video-to-video/) | Wan2.2 VACE + Lotus | 32 / 48 GB | ~143 GB | Transform a basic 3D render into stylized video — depth extracted automatically | | 08 | [Image to 3D](workflows/08-image-to-3d/) | Trellis2 | 24 / 32 GB | ~20 GB | Convert a 2D reference into a textured 3D model with PBR materials |

Bonus Modules

| # | Workflow | Key Model(s) | Min. Rec. Windows / Linux VRAM | Disk Space | What It Does | |---|----------|------------|------|--------|-------------| | bonus-a | [Texture Extraction](workflows/bonus-a-texture-extraction/) | Qwen Image Edit 2511 + Texture LoRA | 24 / 32 GB | ~60 GB | Extract seamless tileable textures from any image | | bonus-b | [Texture → PBR](workflows/bonus-b-texture-to-pbr/) | Lotus + Marigold | 24 / 32 GB | ~10 GB | Generate a full PBR material set (Normal, Height, Albedo, Roughness, Metallic) |

> VRAM — Windows / Linux. On Windows, NVIDIA weight streaming offloads inactive model layers to system RAM. On Linux, the full model must fit in VRAM. > Disk — Per-module figures assume that module installed alone. Many modules share large models (Qwen 41 GB base, encoders); installing all 12 together costs ~450 GB, not the sum of individual figures.

---

Gated models

Two modules require accepting a license agreement on HuggingFace before the installer can download them. The installer will prompt you to confirm you've accepted their terms:

• Module 05 (Flux.1-dev — Black Forest Labs license)
• Module 08 (DINOv3 — Meta license, Windows only).

---

##…