Generate Magnific Level Details and Upscale using Automatic1111

by | Apr 10, 2024 | Resources | 0 comments

This technique was share by Philz1337x and he was the one who came up with this method to generate Magific level details using Automatic1111. He published all of his code as open source but that runs on something called COGS which I don’t know yet how to run on its own. So the interim solution is to use Automatic1111 along with Tile Diffusion and ControlNet.

This method is another great alternative to Magnific which is a paid service and is complementary to another method using ComfyUI I have posted before.

What you need?

The models is recommended but you can also swap and change to another preferred model and similarly add additional LoRA to your workflow. I have tried with a custom model ETV8 which is a paid model that I’ve invested in.

Settings in Automatic1111

In order to start upscaling and adding incredible amount of details in your images you should follow the steps below. There are a few steps so don’t rush and simply take your time, it took me 2 runs to get my head around it. Now I am comfortable exploring and changing them to suit my needs.

  • Start in the Img2img tab and enter the following prompts:
    • Positive: masterpiece, best quality, highres, <lora:more_details:0.5> <lora:SDXLrender_v2.0:1>
    • Negative: (worst quality, low quality, normal quality:2) JuggernautNegative-neg
  • Sampling method should be set to DPM++3M SDE Karras (start with this and they you can experiment with others)
  • Sampling steps: 18 (I used 20)
  • CFG Scale: 6 (start with this and then experiment – this controls the dynamic nature of the upscale and details)
  • Denoise Strength: 0.35 (You can go from 0.3 – 0.9 – this is equivalent to the creativity slider in Magnific)
  • Seed: <set a fixed seed> in below its set to 1337 which I am guessing the number from the author’s handle.
  • Enable Tiled Diffusion
  • Enable Keep input image size
  • Set Method to MultiDiffusion
  • Latent tile width to 112 px and tile height to 144 px
  • Latent tile overlap 4px
  • Latent tile batch size 8
  • Upscaler: 4x-UltraSharp
  • Scale Factor: 2

You can fine tune these settings like the tile width/height and overlap based on the original image size but I suggest you start with these suggested/recommended settings and experiment with them further.

  • Enable Tiled VAE – this will prevent CUDA out of memory error which occurs when you don’t
  • Enable Move VAE to GPU (if possible)
  • Leave most settings to default as shown below
  • Enable ControlNet
  • Enable Pixel Perfect
  • Preprocessor: tile_sample
  • Model: control_v11f1e_sd15_tile
  • Control Weight: 0.6 (you can try 0.3 to 1.6 – this control is similar to the Resemblance slider in Magnific)
  • Rest of the settings are pretty much default

Simply click on Generate and wait a bit for the results to blow your mind. This method using Tiled Diffusion in Automatic1111 will slice the images up, upscale and defuse them further while enhancing the details and then stitch it back up. So far I have tested this process on Cloud GPU in RunPod.io where I provisioned 24GB and 48GB VRAM, and later tested with 16GB, so far I’ve been able to run this successfully on all three instances. Do let me know if you have lower VRAM, and this setup is able to run on this.

The Tiled VAE option will help manage the VRAM more efficiently so make sure you enable this otherwise you may run into CUDA out of memory error.

Upscales achieved with this method

Have a look at the before and after as I share the different ways you can use this method.

We start with a close up shot generated that is upscaled using the recommended model. The original image is upscaled to twice its size with the same settings as setup above. The details are quite high and the resulting image is quite sharp, the hair are well defined.

In this more cinematic looking image, you can see the image is nicely enhanced and added detailed. The main thing is that the resemblance is maintained and no artificial artefacts are added to the image.

A woman in Indian traditional attire, I was curious to see what the results look like. And you can see the details are very nicely and precisely added during upscale. The veil has intricate details added and there are some artefacts appearing on her face but something that can easily be fixed using Photoshop or In-painting in Automatic1111.

Moving into some non human items I took an image of a robot that I had generated with ETV8 and so I decided to upscale with the same model. That’s the only thing I changed to upscale this by 2x again using Automatic1111. Of course you can take the image further by upscaling again.

Here is the resulting image that is now 4x the original image and is compared with the original image. The image does start to appear a bit HDRish in this case so for the second pass one might need to tone down the settings.

Reverting back to an old image that I had sitting around on my laptop, I ran it through Automatic1111 which resulted in quite a detailed image, like you would expect out of Magnific.

Now we explore more images that are upscaled using the ETV8 model (not open source) but available for purchase. This model produces very nice textures and details when prompted correctly. In my case I ran some of my older prompts and their resulting images through this upscale process to see what kind of details are generated.

This first image, I ran through twice the upscale process. That was 2x the original size and ran at 1x on the 2nd pass, this second pass simply adds details to the image without upscaling it. Have a close and careful look at the image, the difference between before and after is quite incredible.

Another example of an image just upscaled from original size to 2x. Wow the face and eyes come out really nice in the end with Automatic1111. The geometric lines and curves are also more well defined.

Here is a pixar style image generated using ETV8 model in Automatic1111, which is now upscaled with greater details using this method.

This final image is a simple 1x upscale run using this method and in this case it acts like a simple detail enhancer without increasing the image size. All this is possible thanks to Philz1337x who shared his method and code publicly for the community. Make sure you check out his other work and follow him.

You may also like:

Flux Dev Artist Study
Ultimate Collection of Flux Dev LoRAs
LoRA Caption Workflow for ComfyUI
Flux Dev Image to Prompt Workflow with Upscaler
Cheap Solution for running Flux Dev using Runpod
Basic Camera Motion LoRA for AnimateDiff

If you'd like to support our site please consider buying us a Ko-fi, grab a product or subscribe. Need a faster GPU, get access to fastest GPUs for less than $1 per hour with RunPod.io

Submit a Comment

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.