Part of my Simple Group. I will also have a plug & Play Module available
Instagram: https://www.instagram.com/synth.studio.models/
Buy me a☕ https://ko-fi.com/lonecatone
This represents many hours of work that I give away for free. If you enjoy it, Please make sure to 👍like and post to my workflows. That is how you support me!!! Also, feel free to ⚡tip 😉
Thanks to ukr8b3g for these nodes
This cool set of nodes uses BBOX technology to create a .json prompt.
Allows 5 regions
easy prompt guides
choose camera angles, backgrounds, styles, and more.
也有中文说明
Description
FAQ
Comments (15)
If I leave ComfyUI-Krea2T-Enhancer (fixed) active, I get visual garbage. I deleted this node and everything works. Thanks.
Hmmm, I'll try it without it.
I am starting to question how good that node is. It seems to mess with words as well
@lonecatone23 There are alternative methods that can even correct lack of emotion.
https://www.reddit.com/r/StableDiffusion/comments/1ukh334/i_extracted_the_values_of_krea_2_safery_filters/
And then, btw: using a rebalance node sometimes results in a black render, so I bypassed it.
@windlike_gust No, That's specifically sage attention that gives you black images
@lonecatone23 Agree. :)
Just before the Clip Text Encoder, I added Text Concatenate (was-ns) for the text prefix and suffix - this helps the positioning system and the overall complement of the composition.
@windlike_gust Can you drop an image with the workflow?
I appreciate the feedback 🙏
@lonecatone23 shure:)
-> https://www.dropbox.com/scl/fi/b1c7xybdffdrsl6maecji/472693118380504.png?rlkey=c6x7n0j7348lgtxjtycq590l4&st=4l1p5gg9&dl=0
This is just a quick sketch, and the prompt is taken from here:
https://civitai.red/images/136209382
I have also disabled the "ComfyUI-Krea2T-Enhancer" node and I no longer get black images either.
@lonecatone23 Two more small clarifications: at this point, according to my A/B testing and, based on the style of the result, I get a more detailed render (you can even see it in the hair) by connect the 4x_NickelbackFS_72000_G upscale model to USDU with the wan_2.1_vae decoder, while the Euler-Ancestral sampler with the Bong-Tangent scheduler is more accurate follows the prompt in the case of two-stage generation with intermediate pixel upscaling. (Of course, if you follow the solution from my second message.)
@windlike_gust Yeah, that'd be great for anime type images, but Nickelback and pixel upscaling trashes realism images. Also, are you generating with regional prompt weights then refining, or are you carrying the weights to both ksmaplers? I can't open your images.
@lonecatone23 In that case, here's json:
https://www.dropbox.com/scl/fi/myozttlj78ld51l43bv7w/472693118380504.json?rlkey=kor9g4pxzw1lxcfarv4nnpuqg&st=bvpubdzw&dl=0
@windlike_gust I'll look at it tonight. Thanks
Oh man, I don't think I've had this much fun with a workflow! Such great work!!! :D
Very interesting node set, it works well