JonXL's Minimax H3 Image Upscaler (LORA + Workflow & Batch)
Consider to download the workflow (from 'optional downloads') for easier image extraction and batch processing.
A conceptual lora and workflow to upscale blurry and pixelated images up to 1920x1920 with Minimax H3.
Extensive testing and multi pass training made it now possible to upscale images with Minimax-H3.
The workflow can take a single image, or take a batch input folder and will automatically upscale image(s) to the nearest given resolution up to 1920x1920.
The upscaler seems to work very well for blurry images taken from a video clip or heavily compressed photos.
(!) Note #0: There are now two LoRas, the 24 Frame Quality lora (1920) and the 1 Frame Speed lora (1536 1 frame) . Each has there speed / quality trade offs. Make sure to download the workflows from the relevant page. (1 Frame workflow requires a custom node, see bottom of the page)
(!) Note #1: This is an upscaler, not a detailer. It will not add more details, but it can extract a lot of detail from blurry and unsharp images. Don't expect it to add detail to already sharp / AI generated content.
(!) Note #2: For maximum quality use the 1920x1920 resolution and the 24 Frame Quality Lora for best results. This can take up to 5 minutes on a RTX5090. The 1 Frame Speed lora however can generate within 30 seconds on a RTX 5090 at the cost of (detailed) quality. Consider lowering your resolution to 1280x1280 for the quality lora, if you have a lower hardware specs or less patience.
(!) Note #3: You can enable the batch mode option in the AIO workflows, supply an input and output folder and it will automatically upscale all images for you, even keeping the file names.
Prompt:
A static still blurry pixelated picture crossfades into a clear sharp detailed picture, 4psc4l3,
// always use this prompt!
Authors comment
Feel free to post your safe for work results on this LoRa's page too!
Please enjoy upscaling your content with this upscaler lora and workflow!
Usage:
Use the following trigger sentence to activate the image generation
A static still blurry pixelated picture crossfades into a clear sharp detailed picture, 4psc4l3,
Seems to work very well on resolutions of 1280x1280 and 1920x1920
It will auto-scale the input image to the nearest resolution
Default values of 24 frames for Quality lora, CFG 1 and Lora strength one should be kept
You can change the CFG and frame count, it might add more detail but it's not recommended unless you like to experiment
For the 1 Frame Speed Workflow, you can use the Turbo Lora to improve contrast at the cost of some consistency, or disable the Turbo Lora for more consistency. I recommend keeping the Turbo lora enabled.
When using the 1 Frame workflows, make sure to clone the custom node and download the special VAE, no turbo lora required.
Technically, you can also use this in Text2Video mode, and generate photorealistic imagery. It was trained on the same dataset as https://civarchive.com/models/2884206/minimax-h3-photorealistic-image-generator-lora-workflow
Let me know your feedback
Known issues:
You may need to install some frequently used nodes, mostly Kijai's nodes and VideoHelperSuite
Description
FAQ
Comments (9)
So I can't just slap this lora on an existing workflow?
Short answer, probably not.
Long answer, it's meant to upscale input existing images to output upscaled images. You kinda need to pick the last frame for it to work. The supplied worklfow is ideal for that.
If you just toss it in a lora loader to generate videos, well, it's not meant for that, but still it will most likely make your output a bit more photoreal but could dampen motion.
As described in the description, it's mostly useful to upscale blurry pixelated unsharp images (like low res single video frames or old photographs), it won't do much for static AI generated content (usually already sharp)
The only reaaon you could add this to another workflow is if that workflow uses an input image loader, and you would intend to do a upscale pass before plugging it in a i2v sampler or ref2v sampler.
Got it. Looking forward to the super simplified workflow you said you were working on elsewhere then.
@NeilHamburger It's already uploaded in the workflow zip, called 'Simple'
@JonXL Oh, I mistakenly thought there was a simpler version coming. The simple one is still pretty complex.
@NeilHamburger Installing the nodes, Loading the models, loading an image and pressing "prompt". It doesnt get less complex in ComfyUI. At some point I am not sure what to tell people :)
Interesting finding.
Consider playing with step count, such as 6 instead of 20 while upscaling with the 24 Frame Lora (1920).
Seems to oversharp less, and saves 66% of the render time.