Upscale & Add Detail Locally With Stable Diffusion (FREE Magnific)

Ready to take your images to the next level of brilliance and detail? Meet Stable Diffusion's MultiDiffusion extension; a free, local enhancement solution that rivals the acclaimed Magnific tool. With MultiDiffusion, you can enhance and upscale your images with stunning precision and clarity, all without breaking the bank. Discover how MultiDiffusion revolutionizes image enhancement, offering unparalleled precision and clarity for adding detail and upscaling your visuals to new heights.

Controlnet + Tile model

The first requirement is Controlnet and the Controlnet Tile model. ControlNet is a must have extension for Stable DIffusion, I recommend watching our tutorial on How to Install ControlNet if you don’t have it installed already!

Tiled Diffusion Extension

Next up is the Tiled Diffusion or sometimes referred to as Multidiffusion extension. To install this extension open up your Automatic1111 interface and go to the extensions tab. From here click on “Install from URL” and paste the following github link: https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111.git Then go ahead and click on “Install.”

When the Tiled Diffusion extension has been installed go to the Installed tab and press on “Check for updates” followed by “Apply and restart UI.”

Before we can start adding details and upscaling our image, we need a starting image. I will quickly create an awesome image with the ZBase-XL checkpoint together with the XL Weapon Spell Rings LoRA. Alternatively you can choose any image you like.

• Prompt: a wolf warrior standing on two legs in a fighting stance, he is wearing medieval samurai armor, he has a magical turquoise aura around him, intricate details, turquoise <lora:XL_Weapon_Ring_Spells_-_By_HailoKnight:0.7> Spell rings
• Negative prompt: cartoon, blurry
• Sampling method: DPM++ 2M Karras
• Sampling steps: 30
• Width & Height: 1024x1024
• CFG Scale: 5
• Seed: 2336269327

This step is optional but I will now run the image with the same seed again with Hires. Fix enabled. For the upscaler I will use 4x-UltraSharp. Set “Upscale by” to 2 and decrease the “Denoising strength” to somewhere between 0,1 and 0.3.

Don’t forget to enable it and use the same seed for the generation. Now hit Generate again.

Once you have your base image ready we can send it to image-to-image by pressing the little painting under the image. If you want to use a different image you can simply drag and drop your image onto the img2img canvas.

img2img Settings

The first thing we need to change is our checkpoint to a SD 1.5 checkpoint, unfortunately Multidiffusion does not work properly with SDXL models yet as it results in the image having a pattern overlay, the creator is working on this so hopefully soon we can use this with SDXL models.

• Checkpoint: Juggernaut (this is a good all round checkpoint to use covering all sorts of styles.)
• Modify prompts by removing descriptive keywords.
• Prompt: hyper detailed, intricate details, extreme quality, 8k, HDR
• Negative prompt: cartoon, painting, illustration, (worst quality, low quality, normal quality:2)
• Sampling method: DPM++ 2M Karras
• Sampling steps : 30
• Denoising strength: low value (between 0.3 and 0.6) to maintain the image or use a higher value to add more details

Tiled Diffusion Settings

Scroll down, open the Tiled Diffusion extension and enable it.

• Change the Method from MultiDiffusion to Mixture of Diffusers, there is not much difference between the two but the creator of the extension recommends Mixture of Diffusers.
• Set the Latent tile width & height to 128, and depending on your VRAM you can increase the Latent tile batch size. With 8GB of VRAM I recommend leaving it at 4.
• For the Upscaler I will use 4x-Ultrasharp, but you can choose anything you like.
• A good pre-installed upscaler is R-ESRGAN 4x.
• Leave “Scale Factor” at 2 for a 2x upscale.

Noise Inversion Settings

Noise inversion is where the magic happens, this will add a lot of details to your image. Enable Noise Inversion and set the “Inversion Steps” to 50. Reducing the “Renoise Strength” adds detail, if you set this all the way to 0 I recommend increasing the Denoising strength to around 0,75. You can play around with these settings as there is no “perfect” setting that will work for everyone, a good starting point is 0,10.

Tiled VAE Settings

Lastly scroll down and enable the Tiled Vae extension. Here we will only enable the “Fast Encoder Color Fix” option. This makes sure our image stays vibrant and doesn’t lose color or will look washed. The Encoder Tile Size is adjusted automatically based on your GPU’s VRAM.

ControlNet Settings

The last thing we need to do is enable ControlNet, scroll down and open the extension. Check the Enable and the Pixel Perfect checkboxes.

Now change the following settings:

• Pixel Perfect: Yes
• Control Type: Tile/Blur
• Preprocessor: tile_resample
• Model: control_v11f1e_sd15_tile
• Control Mode: ControlNet is more important

Now go ahead and click Generate! This may take a minute or 2 depending on the size of your image and the speed of your GPU. The larger your image, the longer the process.

Let’s take a look at the image we just upscaled and enhanced with MultiDiffusion! For this upscale i used a Denoising Strength of 0,5 and Renoise Strength set to 0,1.

Let’s take a look at another image I upscaled and enhanced with MultiDiffusion! For this upscale i used a Denoising Strength of 0,75 and Renoise Strength set to 0!

As you can see, sometimes the extension can add too many details! Play around with the settings and see what works the best for your image. I recommend starting off with a smaller size image so the generation won’t take as long.

If you are satisfied with the amount of detail you can increase your image’s resolution without adding detail too. For this simply disable Noise Inversion and ControlNet. Then lower the denoising strength to 0,1 and generate!

For quicker generation, start off with a smaller image, I showcased how to add details and upscale a 2048x2048 image. Obviously this is quite large already and you don’t have to copy this.

Another way to speed up generation is by reducing the “Upscale By” in the Hires. Fix settings and the “Scale Factor” in the Tiled Diffusion settings. If you don’t want to upscale the image to add detail you can set the Scale Factor to 1, which will dramatically increase your generation speed.

If you want more change in your image you can play with the “Retouch” setting in the Noise Inversion section. However this option is very sensitive, a value of 2 can already change your image a lot. This will multiply if your Renoise Strength is very low.

If your image has too much contrast you can try to lower the CFG Scale. This can reduce contrast and artifacts in your image.

One setting I haven’t mentioned is the “Latent tile overlap” in the Tiled Diffusion section. The default value is 48, but you can decrease this all the way to 8 if you like.

In conclusion, MultiDiffusion emerges as a formidable tool for image enhancement and upscaling, showcasing remarkable precision and detail. With features like Noise Inversion, it empowers users to infuse images with intricate details, elevating resolutions while maintaining complete control.

This extension opens doors to creativity, allowing for the enhancement of images with unprecedented levels of detail. With Multidiffusion, users can explore new dimensions of expression, pushing the boundaries of visual storytelling and creativity.