Real-Time Image Generation from Scribbles with ComfyUI

akuma.ai has been getting a lot of attention lately (apologies if this info is out of date). Looking at what it does, I felt something similar should be reproducible in ComfyUI — so I gave it a try.

Workflow File

live_painting.json
drive.google.com

Setup Procedure

1. Set Up ComfyUI

Install ComfyUI using the following method.

【Stable Diffusion】ComfyUIを使って画像生成AIで遊んでみよう【導入編】

>-

blog.otama-playground.com

2. Install Extension

Install the following extension into ComfyUI. Using ComfyUI-Manager makes this easy.

GitHub - AlekPet/ComfyUI_Custom_Nodes_AlekPet: Custom nodes that extend the capabilities of Comfyui

Custom nodes that extend the capabilities of Comfyui - AlekPet/ComfyUI_Custom_Nodes_AlekPet

github.com

3. Download and Place the Model

Download LCM-LoRA and place it in the LoRA directory.

LCM-LoRA Weights - Stable Diffusion Acceleration Module - LCM for SD 1.5 | Stable Diffusion 1.x LoRA | Civitai

LCM-LoRA - Acceleration Module! Tested with ComfyUI, although I hear it's working with Auto1111 now! Step 1) Download LoRA Step 2) Add LoRA alongsi...

civitai.com

4. Import the Workflow

Download from the link below and import it into ComfyUI.

live_painting.json
drive.google.com

5. Configure Models

Set the LCM-LoRA in the Load LoRA node, and set your preferred base model in the Load Checkpoint node.

6. Configure Auto Queue

Set things up so that generation runs automatically whenever the paint canvas or parameters are updated.

Specifically, configure Extra Options as shown below.

Setting of Auto Queue
Setting of Auto Queue

7. Generate

Once you click Queue Prompt, your scribble will be reflected in the Preview Image node almost in real time.

Tips

  • Adjust the prompt to match what you want to generate. Using an embedding like EasyNegative in the negative prompt can help improve results.
  • The seed is fixed so the output doesn’t change randomly between generations. You can update it manually via the Primitive node if the current seed isn’t producing good results.
  • LCM-LoRA isn’t the only option for real-time generation. Lightning, Turbo, and Hyper models can all work as replacements.

Generation Result

Apologies for my poor drawing skills…

Scribble
Scribble
flat2dAnimerge
flat2dAnimerge
cardosanime
cardosanime
mistoonanime
mistoonanime
darksushimixmix
darksushimixmix
qteamix
qteamix
dreamshaper
dreamshaper

Conclusion

The results look pretty decent, don’t they? Without a video it’s hard to fully show it off, but if you actually try it, what you paint gets reflected nearly in real time — which makes it genuinely fun to play with.

Being able to run something like this for free locally is exactly the kind of thing open source excels at. If this looks interesting to you, definitely give it a try.