Realtime AI Generation Local Install Guide (AI Drawing) for FREE on your PC!
Based on MattVidPro's video on YouTube. If you like this content, support the original creators by watching, liking and subscribing to their content.
Check Task Manager → Performance → GPU for dedicated GPU memory; shared-memory-only GPUs are treated as incompatible.
Briefing
Local, real-time AI drawing can run privately and for free on a Windows PC—provided the machine has an NVIDIA GPU with dedicated memory (at least 8 GB recommended). The setup hinges on installing ComfyUI plus ComfyUI Manager, downloading the “SDXL Turbo 1.0 do saafe tensors” model, optionally adding upscale models, and then using a custom ComfyUI workflow that enables prompt editing and image generation to respond immediately as the user draws.
Hardware checks come first. On Windows 10 or 11, the process starts in Task Manager under the Performance tab, then the GPU section. The key requirement is dedicated GPU memory: integrated GPUs with only shared memory won’t work for this guide. The tutorial frames 8 GB dedicated VRAM as the bare minimum for “real-time” behavior, while the example system—an AORUS 17X laptop with an RTX 4080—has 12 GB dedicated VRAM, which is treated as comfortably sufficient.
Software installation is then staged around four downloads and two utilities. Seven-Zip is recommended to speed up unpacking large archives. The guide instructs users to download SDXL Turbo 1.0 do saafe tensors (about 14 GB), ComfyUI (about 1.4 GB), and optional upscale models from a Google Drive link. Git is installed next because ComfyUI Manager relies on it. After that, ComfyUI Manager is installed using a script copied from the instructions, and the user is told to wait for all downloads (especially the large SDXL Turbo file) to finish before continuing.
Once ComfyUI is extracted to a chosen folder, the workflow requires placing the SDXL Turbo checkpoint into ComfyUI’s checkpoints directory and copying upscale model files into the upscale models directory. A custom workflow file (“Matt vidpro realtime drawing final. JSO” from a Google Drive download) is then loaded via ComfyUI Manager. The guide notes that certain “no time were not found” and missing-tab warnings may appear during graph loading, but it treats them as expected. The critical step is clicking “Install missing custom nodes,” installing two required nodes (with a restart after the second), and confirming the interface loads cleanly.
With the system running, the interaction loop is built around prompt queuing and live editing. Users press Q prompt to generate from a prompt like “photo of a tabby cat,” then use “extra options” to set batch count to one and enable Auto Q so prompts update automatically. Editing the prompt text changes the output in near real time. For drawing, the workflow opens an “edit in another tab” drawing canvas; releasing the pen triggers image updates. The AI responds to scribbles, but the tutorial repeatedly warns that the prompt must reflect what was drawn (e.g., specifying “photo of the dog to the left” when scribbles imply a dog). The user demonstrates changing composition, perspective, and style by drawing and rewriting prompt details on the fly.
After a satisfactory result, upscaling is performed by disabling Auto Q, reloading upscale modules, and queueing the upscale models (example preference: “4X NM KD super scale”). Outputs are saved under ComfyUI’s output folder. The guide closes by noting that exiting involves closing the ComfyUI windows and rerunning “run Nvidia GPU,” and it highlights the sponsor’s AORUS 17X as a machine capable of handling both recording workloads and local real-time AI generation.
Cornell Notes
The guide lays out a Windows-first method to run local, real-time AI drawing using ComfyUI with a custom workflow and the SDXL Turbo 1.0 do saafe tensors model. The most important prerequisite is NVIDIA hardware with dedicated GPU memory—8 GB dedicated VRAM is treated as the minimum for real-time performance, while shared-memory-only GPUs won’t work. After installing Seven-Zip, Git, ComfyUI, and ComfyUI Manager, users place the SDXL Turbo checkpoint into the checkpoints folder and upscale models into the upscale models folder. The workflow then enables Auto Q prompt updates and a drawing canvas where releasing the pen triggers immediate image regeneration. Final images can be upscaled (with Auto Q disabled) and saved to ComfyUI’s output directory.
What hardware requirement determines whether this local real-time drawing setup will work?
Why are Seven-Zip and Git installed early in the process?
How do the SDXL Turbo model and upscale models get wired into ComfyUI?
What makes prompt changes and drawing updates happen in real time?
What’s the workflow for upscaling after a drawing is complete?
Review Questions
- What specific Task Manager GPU metric does the guide use to decide whether the setup will run, and what happens if the GPU only has shared memory?
- Where do SDXL Turbo 1.0 do saafe tensors and the upscale model files need to be placed inside ComfyUI?
- How do Auto Q and the drawing pen release event work together to produce real-time updates?
Key Points
- 1
Check Task Manager → Performance → GPU for dedicated GPU memory; shared-memory-only GPUs are treated as incompatible.
- 2
Plan for at least 8 GB dedicated VRAM for real-time performance; the guide’s example system uses 12 GB dedicated VRAM.
- 3
Install Seven-Zip for faster extraction, then install Git because ComfyUI Manager requires it.
- 4
Download and place SDXL Turbo 1.0 do saafe tensors into ComfyUI’s checkpoints folder and upscale models into the upscale models folder.
- 5
Load the custom “Matt vidpro realtime drawing final. JSO” workflow, then use ComfyUI Manager’s “Install missing custom nodes” and restart after the second node.
- 6
Use Q prompt plus Auto Q (batch count set to one) for automatic prompt updates; use the drawing canvas where releasing the pen triggers image regeneration.
- 7
Disable Auto Q to upscale, reload upscale modules, queue the upscaler, and retrieve saved outputs from ComfyUI’s output folder.