LTX 2.3 - LTX Image Upscale (LORA + Workflow)

Make sure to download the *updated* workflow (V3) from the 'required downloads' as well!

There is an additional downloadable lora called 'depixelator' that improves the upscaling even more. (click on the gray download panel to make it expand). There is also a 'lite' version (scaled_0.5) of the upscaler lora in case you get oversharped results.

A conceptual lora and workflow to upscale unsharp, blurry low-resolution and slightly pixelated images to high resolution images up to 1920x1920.

The workflow captures the last frame out of a fade-in video (lora trained) where the upscaling takes place.

Usage:

• The workflow is heavily recommended

  • (As it has extremely specific parameters)

• Preferably, use the following sentence to activate the image generation (no additional words required)

  • 4psc4l1fy, a blurry pixelated image transitions into a static sharp high quality image

• The initial parameters are a good start, but if you need more extreme upscaling then try increasing the CFG and Upscalify Lora strength.

• Please use 'Megapixel' mode, and use megapixels between 2.06 and 3.68.

  • Try to estimate the megapixel value to use, closest to the following input image aspect ratios:

    • For 16:9 or 9:16 use a maximum of 2.06 megapixels

    • For 4:3 or 3:4 use a maximum of 2.75 megapixels

    • For square images (1:1) use a maximum of up to 3.68 megapixels

  • If you are uncertain stick to 2.06 megapixel. However, trying to match the megapixel amount (especially if its closer to a square instead of a wide rectangle) will make your images get sharper!ĺ

  • Another way of working out the megapixel value is seeing what thel longest size of the output image is. Make sure it does not exceed 1920 pixels for the longest side.

• The use of a pre-blur pass is also included, adding blur can ironically help with adding minor detail.

  • Default value is 5, but you can set it to 0 or higher if you want to experiment

• Feel free to post your results, though keep it decent

• Let me know your feedback, as this is the only image upscaler using LTX in this manner that I know exists for LTX so far :)

Known issues:

• This workflow will upscale unsharp blurred and somewhat pixelated images very well

  • Note however that it will not automagically fix plastic skin, it's not meant as a detailer.

    • (You could however try to add film grain or other noise to the input image, to experiment with it's effect on the output using the upscaler)

• You may need to either install Kijai's nodes, or bypass the ChunkFeedForward node inside the subgraph, follow the install instructions here:

  • https://github.com/kijai/ComfyUI-KJNodes