Amazing controlnet! Thank you!
Came by to say thank you to the @jasperAITeam for creating and releasing this controlnet! I've creating a ComfyUI workflow around this and I've been having a lot of fun using it and really impressed with the quality.
You can grab my ComfyUI workflow on my Github: https://github.com/dicksondickson/dickson-sci-fi-enhance-upscale . The "promax" versions are they ones that uses JasperAI controlnet and Flux.
Please keep up the amazing work!
Tried 4K Pro Max and 4K Advanced - the Flux version clears all the imperfections MUCH better, but it really distorts the faces - especially the eyes! The eyes become unrecognizable! So, until it's fixed - 4K Advanced is the better version.
Thank you for trying out my workflows! Yes, Flux smooths things out too much - I think a lora needs to used to add skin texture back in. My 4K advanced and 8k advanced versions (which uses SDXL) is more mature and retains skin texture much better.
No problem bro) The biggest problem with your flux workflow is the eyes, but i believe, they are screwed before the flux - at the SUPIR step.
Try using Realvis XL 4.0 for the SUPIR step - seems like Boltning HyperD is giving weird issues sometimes. Some people told me they are getting better results with that setup.
@dicksondickson I tried out your workflows and they work well for images that has some detail.
I notice on the Demo version here is able to add allot of detail to images that has hardly any. Were you able to replicate a similar output
Demo: https://huggingface.co/spaces/jasperai/Flux.1-dev-Controlnet-Upscaler
( Seems to be broken at present )
Just tried using Realvis XL 4.0 for the SUPIR step - it made the picture MUCH worse! The faces (especially the eyes) are unrecognizable!
@Tequiila Glad you like it. I'm thinking of testing out noise injection to get skin details and other details back in but I haven't found the time to work on it yet. Try it out and let me know your findings!
@Johndoe6801010 Can you post your source image? I've had good success with a wide variety of images but sometimes the prompting needs to be specific for the SUPIR step.
@dicksondickson Appreciate you setting that up. very helpful.
This is the result from the demo.
The only way I can get a similar result is using controlnets (SDXL). I not sure if something like this was setup for the Flux demo. (I'm guessing: not)
@Tequiila In my workflow, I recommend you play with the controlnet strenght and the denoise settings. Also, try my SDXL based upscale workflows (4k/8k Advanced) on my github. That might give you better results than my Flux based ones for now. Would love to see your results.