ComfyUI: Tiled Diffusion upscaling with Zimage Turbo

Step-by-step

1. Open workflow 3 in ComfyUI.
2. Install the 'Tiled Diffusion' node via the Custom Nodes Manager if not already present.
3. Load your image using the Load Image node.
4. Resize the image down to 1 megapixel to prevent VRAM crashes.
5. Apply the Upscale Model node with a factor of 0.5 to achieve a 2x upscale.
6. Connect the upscaled image to the K Sampler using the Zimage Turbo model.
7. Configure the Tiled Diffusion node to manage image processing in smaller segments.
8. Adjust the Denoise value (e.g., 0.05 to 0.4) to control the level of detail and creativity.
9. Use the VAE Decode Tiled node to output the final image.

Options

• Increase the megapixel resize value if you have high VRAM.
• Reduce the tile batch to 1 for low VRAM systems.
• Reduce tile width and height to 512 pixels for better memory management.
• Use a normal VAE Decode instead of Tiled VAE Decode if VRAM allows.
• Bypass the Tiled Diffusion node to process the entire image at once if you have sufficient VRAM.

Watch out for

• The Zimage Turbo and upscale models must be downloaded into the correct folders before starting.
• Setting the Denoise value too high can cause visible tile seams or strange textures/errors in the image.
• Images that are too large at the start of the workflow may cause the system to crash.

Tips

• Stick with the default 1 megapixel resize for stability.
• Describe your image in the prompt to get better results, though simple prompts often work.
• Run multiple seeds and adjust Denoise values to find the best balance of detail and accuracy.
• Use a lower Denoise value (around 0.05) to keep the result more similar to the original image.