ZebraDraw: localLLM Bulk Drawing Orchestrator

ZebraDraw is a bulk drawing tool, that can allow a localLLM to draw hundreds of batched photos all night!

Licensing: Please ensure you have the correct license for whatever diffusion engine you are using if you are going for commercial work.  ZebraDraw is completely opensource

Pull and run it as a Docker Image

• This can be easily pulled and ran with a few simple commands.

sudo mkdir -p \
  /opt/zebra-draw/models \
  /opt/zebra-draw/config \
  /opt/zebra-draw/results

docker run -d \
  --name zebra-draw \
  --gpus all \
  --restart unless-stopped \
  -p 0.0.0.0:8080:8080 \
  -p 0.0.0.0:8081:8081 \
  -p 0.0.0.0:8082:8082 \
  -e MODEL_PATHS=/models \
  -v /opt/zebra-draw/models:/models:ro \
  -v /opt/zebra-draw/config:/config \
  -v /opt/zebra-draw/results:/results \
  cnmcdee/zebradraw:latest

You will probably want to give your local user access to those directories it's just easier for testing so chown -R user:user /opt/zebra-draw

Port 8081 - GPU Diffusion Model Setup.

We needed to show this port first, as port 8080 and 8082 will not work if this is not setup. There are lots of different ways this can work but simply you will want a model configuration that looks like this:  Whatever the local ip of your machine is you should be able to find it by opening a browser window: 192.168.0.<your ip>:8081 which you type into your local browser.

There are three parts to this model, not all are sometimes required..

• Main image model krea2_turbo_Q6K - 9.86 - It draws.
• VAE wan_2.1_vae - 242.0 mb The translator between latent space and normal engine space.
• Text Encoder Qwen3VL-4B-Instruct-Q4-K_M This converts your text into image embeddings that your wan_2.1_vae will require.
• Click off Standalone diffusion model if you are using this setup.

To get and pull these models - remember to save them to your /opt/zebra-draw/models directory.

cd /model

wget -c -O krea2_turbo-Q6_K.gguf \
"https://huggingface.co/vantagewithai/Krea-2-Turbo-GGUF/resolve/main/krea2_turbo-Q6_K.gguf"

wget -c -O qwen_image_vae.safetensors \
"https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/vae/qwen_image_vae.safetensors"

wget -c -O Qwen3VL-4B-Instruct-Q4_K_M.gguf \
"https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3VL-4B-Instruct-Q4_K_M.gguf"

If it is loaded cleanly - you will will not get an error, it will look something like this at the bottom inside your sd-server.log

Port 8080 - Single Image Creation

• Once you are comfortable that your GPU inference engine is setup correctly, you can test a single draw, this is what Port 8080 is for: 192.168.1.<your ip>:8080 in your local browser
• Seed -1 will make it completely random thus the same prompt will draw exactly the same if you used any number other than -1
• Steps 20 is the number of iterations  that the LLM will take to ensure empty spaces are filled int
• CFG Scale  Low 0-2 get really creative 5-7 be pretty strict about following the prompt. NOTE: After some investigation it is recommended in this instance to set it to 0.0

Note: Your diffusion model may not support resolutions very high - it is important to test it here.

Port 8082 - Bulk Drawing Images

• 192.168.1.<your ip>:8082 in your browser
• It is important to trial this on one or two images before bulking. The interface will look like this:

As it draws each one - it will show the image in the progress window, you can use the DOWNLOAD ALL button to have it pack and send you a zip file

Use Chatgpt.com to write up 100 drawing prompts, here is an example prompt list:

Just drag-n-drop this file into the drag-n-drop box on the port and upload it!

Conclusions

This is really powerful stuff - it is local and it is completely opensource and free. You can mix-n-match different models that you can save. It runs in a docker.  If you need to see what your docker container looks like while running: docker ps -a, and finally watch nvidia-smi will show you a hard working GPU:

Source Included