Crystal Clear XL

@WarAnakin thank you 4 this info

yes

I bet there is. But yeah not by us :)
Pretty sweet images.

Hi, NOOBDA! I'm from the Crystal Clear team.

What do you use to generate your images? I personally use SD.Next: https://github.com/vladmandic/automatic

The instructions for SDXL are here: https://github.com/vladmandic/automatic/wiki/SD-XL

Please also try upgrading the installation of whatever tool you're using.

I use Automatic1111. Have installed SDXL 1.0 base and refiner and Crystal Clear too. Also using the recommended VAE for Crystal Clear.

The VAE listed here is the same as the official one it seems. They have the same hash.

@jbluew i will keep trying

Same, I get a garbeled mess

@TommyAI I've created hundreds of images with this model and have seen mostly outstanding images coming from it. What do you mean by "garbled mess"? Are you using the refiner? If so I suggest not using the refiner. This model really doesn't need it, IMO. :)

@NOOBDA Feel free to message me on Discord (same username), I'll help you out there.

@chamois_jubel Hi. I figured out how to work. Thanks!

Wait for new A1111 update, or for now use ComfyUI or Fooocus, it works with only 16GB of System RAM and 6GB of GPU VRAM.

Fooocus rocks! And this model also, it's lightning fast towards the base SDXL 1.0

runnin this on 4gb vram on comfy

It's the first time I saw mentioning of Fooocus in this thread. Quickly looked at the description and this part (in the List of "Hidden" Tricks) #8 "Separated prompts for two different text encoders seem unnecessary... we refrain from implement this." I didn't like at all. Just today I decided to test what's going on with G\L stuff and although I have no proof, my feeling is that it does matter what and in what combination you put into the two encoders. And Comfy gives great flexibility of what you can do and how, so I'd also recommend using Comfy. Especially after seeing people claim it works with 4gb for SDXL.

@green_anger yes, the G/L is new SDXL features. ComfyUI also have this official node (not a custom node) for G/L prompt, see number 5 from here https://huggingface.co/datasets/marhensa/comfyui-workflow for the implementation. G Prompt, is for Linguistic, like normal spoken word. L Prompt, is term / separated word, just like what it used to be on regular SD as before. Also see here for more depth information https://redd.it/15ggn9w

@marhensa Thanks! I'd seen the mini study you linked, but at the time I wasn't familiar with ComfyUI that much to try. Before the small test I mentioned, I also found a short note about the difference between G and L and it was like you said, G for linguistic and L for prompt salad. So, I created two inputs in Comfy, let's call them G and L. The best results (to my taste) were when text_g received G and text_l G + L (concatenated with another node). In the "L" I put only short guidance words like "RAW photo" or "high res", everything else goes to the "G" in a spoken style. Would be interesting to learn if it makes any difference for the negative prompt.

@green_anger Thank you! I tried your method, and yes, it's better somehow. 'G' goes to 'text-g', and the concatenated 'G+L' goes to 'text-l'. If 'text-l' only gets the text from 'L' prompt, it doesn't seem to do much. Thanks for sharing. I'm currently concatenating manually by copy-pasting. Would you mind sharing the concatenation custom node name?

@marhensa Sure thing. It's from this pack - https://civitai.com/models/33192/comfyui-impact-pack - called "Text Concatenate", it accepts 4 elements, if you need more, just nest them. The same author created this manager - https://civitai.com/models/71980/comfyui-manager - if you haven't tried it, I highly recommend, it automates a few things. ComfyUI is great for automating things (copy-pate is getting really annoying after a while), the biggest thing I miss is the on/off switch, right now you must reconnect things, but I think it's a limitation coming from the JS UI library (litegraph.js if I'm not mistaken).

@green_anger oh, I already installed those two custom nodes, I use them for another purpose, thanks for pointing that out. yeah, I agree, ComfyUI is something scary at first, but it's helpful for many cases. I am now familiar with ComfyUI, even I wrote a custom node too while learns Python, it's available in ComfyUI Manager, named "Recommended Resolution Calculator" if you want to check it out, and here is the github page for that: https://github.com/marhensa/sdxl-recommended-res-calc . About switch part, I personaly use connect-disconnect thing with reroute node, but I think the new ComfyUI already implement right click nodes > bypass to disable some nodes.

@marhensa I could have seen this one while was learning about available nodes but it was above my current knowledge :) I may return to it when I get better understanding of the latent space. Apparently it's not possible to send DM in civitai, so I simply started your repo and followed you here. Publish your node in here, people do that, see if it gets traction. Btw, for the image size I use this tiny node - https://github.com/wsippel/comfyui_ws - it's very handy, don't need to remember all the combinations and it's according to the official paper.

@WarAnakin I don't use Discord. Could you please just let us know if images made with this model can or can't be used for commercial purposes? It's pretty vague right now and not very clear on the available uses. I've never heard of a Stable Diffusion model creator not letting the users use the output images for commercial purposes. Please just yes or no - can the images generated with it be used commercially?

@WarAnakin Thank you for taking the time to respond and clarifying! It's very much appreciated! If I end up using any of the images I make with your model for any commercial purposes I will send you a donation via the link on this page. I appreciate you letting me know! :)

When they say, flagging their content, I imagine they are talking about adding a digital signature to all renders generated through this model. They may be doing that already, however I believe they can't claim copyright since this is built on stability.ai's copywritten work and must follow that open source licensing.

@WarAnakin a1111 with civitai shortcut to do the downloading...

@WarAnakin sorry, forgot to let you know, I redownloaded it with civitai shortcut, and this time it actually worked. Apparently there was one of those dumb "connection errored out" situations during the dl that caused it to miss some data the first time through.

I'lll generate something and post it to the page. :)

@WarAnakin Stunned... just absolutely stunned by the results!

@WarAnakin I tried many XL models for this contest and CCXL was definitely the best fit for my purposes (and still is)
Hope to see some new even better version soon, maybe with a bit of artistic accent? ;)

You should check out https://sd-parseq.web.app/ to drive deforum.
It can be used to sync audio tracks to the deforum controls getting some awesome results in time with music.

@b2d2 Cool. I'll check it out if I decide to do another of these. Thanks.

I loved the video, especially the subtle circular rotating effect around the one minute mark.

How many images did this take?

That would be an interesting fact to post on your channel, just to show the sheer amount of hours and work involved to produce a video lasting a few minutes.

Check out RealVisXL 2.0 too, I was really happy with the realistic skin it produced. It mixes this model into it. No longer do I have to go into Lightroom and "shittify" my images to get rid of the airbrushed skin 😂

Pretty please? ♥

@WarAnakin Oh yea, I've used it a lot, merged it into a few checkpoints... Sadly, I haven't used lightning for a while. It seems like everybody has moved away from Lightning and Turbo :( (I really like lightning, I get much better images w/ it than SDXL, but I'm using SDXL & Pony currently)

Did you test against the updated model, Crystal Clear One? https://civitai.com/models/415107/crystal-clear-onevs1

@koitenshin yeah I tried. And results are fugly. Mushy faces and background. My (All done with OneTrainer, Fine Tuning base SDXL model). LoRAs work great with this Crystal version, but look like shit really with the newer. Thanks for the link though ;) I totally missed this One_vS1 version ;)