Hey folks I added another workflow (see description) if you want to download the models manually. I've read on other socials and here on YT some people can't use the node that automatically downloads the models for some reason. Install directions are on a note window in the workflow.
That's not a bad idea and fairly easy to do if you don't mind the extra step but if you are familiar with ComfyUI it's pretty simple to put together. I'll be playing with img2img and trying out the IP-Adapter to see how it is the next few days. Thanks for confirming, good to know it's not just me. I hope they plan on fixing the artifacts though. Kind of a deal breaker for long term use.
Apprently there are still concerns for enterprise license. Here is a good video that breaks it down th-cam.com/video/23tEjvCNZHA/w-d-xo.htmlsi=BGAmCeGL4jt02vEz
@@cyberhard well keep in mind this is a fully trained model and Auraflow is still in beta and not fully trained. Prompt coherency is better with Auraflow I’ve found so far but again, different styled base models and both are fun to use. Will be very interesting these next few months
Yes finally!, although i think auraflow has a better promp adherence,i would say the photorrealism in kolors is better, now you could try hunyuan😝. I liked that You mention about celebrities, this new models seems to not be a big fan of them sadly
hahaha yeah that's another one I can do a video on. There is even Ella which looks interesting. The celebrities are a bummer but I did notice it knows Asian celebrities. hehehe. It's all good and fun for me, kind of a nice change from SDXL and while we wait for the updated SD3 medium.
@@MonzonMedia then My theory was right lol, of course it knows about asian ones haha, sadly i don't know any beside Lucy Liu xD, and after that ,You could try 900m pixart to SEE how it goes 👌🏻
Already recorded a video on Pixart Sigma and the 900M variant but I didn't like how it turned out. Going to do something else and re-record. Pixart is my least liked model out of everything I've tried lately but I still enjoy using it.
I really want to try this model but the auto download won't work for some reason.. I've tried manually downloading but the workflow errors out and says it can't find the model.. I tried placing them in the suggested folders on the github page but no luck..
Strange, works for me every time. Keep in mind it can take a few minutes to download the models. Also, you can check your folders of the models are there, it's located in a different place than usual. For me my address is like this, yours should be similar. Note it should be under diffusers if you do the auto download node. C:\Users\ermin\comfyui\app\models\diffusers\Kolors\unet. In the same Kolors folder there is a folder for the text encoder, you should have a few models in there as well.
AuraFlow значительно уступает Kolors. Я тестировал обе. Kolors на текущий момент - одна из лучших генеративных моделей для SD. Но пока есть одна проблема - с ним не работает InstantID, но это, вероятно, скоро починят.
Well keep in mind Auraflow is only in beta at the moment and it is known that it requires further training. The prompt adherence is also a bit better than Kolors and close to SD3 coherency. That being said, I really like Kolors considering it's a base model BUT the artifacts is a major flaw in my opinion. I hope they release an updated model that corrects that.
@@MonzonMedia Согласен, что артефакты - это проблема, но будем надеяться на лучшее. Сейчас это спокойно решается hires. И да, претензий к бете нет. Ждем новых моделей, ведь в каждой можно найти какую-то прелесть :)
Very true! I'm having fun using these new models as you said each one has a unique style and to be honest I'm kind of bored with SDXL models as they are all very similar.
As far as I know only ComfyUI at the moment, there is an online demo where you can try it. huggingface.co/spaces/gokaygokay/Kolors. I haven't searched around for the latest posts yet but will see if there is any news on support for other platforms.
I suppose it's possible however the artifacts you see are likely due to low quality images, synthetic images in the data set. I'm kind of surprised that they would release this way. I hope they plan on improving it.
Actually MZ is the better version, you get access to more Samplers options, The only problem I'm noticing with Kolors, is it doesn't like to do full body humans, maybe I'm prompting wrong
True, haven't gotten too far with testing other samplers (and nodes) as I just started playing around with Kolors these past few days. Will be fun to discover it's capabilities. As for full body, always helps to prompt things more specifically, like "standing on the side walk" or describing pants, shoes, ground etc. Full body isn't the best word to use and sometimes will give you a "full figured" person 😊
I often have more success with 'full body view'. Also interesting is 'full body view from side' or 'from above' etc. However, I have not tested this with Kolors yet. 🙂
Hey folks I added another workflow (see description) if you want to download the models manually. I've read on other socials and here on YT some people can't use the node that automatically downloads the models for some reason. Install directions are on a note window in the workflow.
Have you tried Kolors yet and are you also experiencing the weird artificating? If there is something I'm missing please feel free to let me know.
Thank you another great tutorial, Kolors is my new favorite. I will try refiner too, previously I upscaled with Fooocus to get rid of the artefacts.
That's not a bad idea and fairly easy to do if you don't mind the extra step but if you are familiar with ComfyUI it's pretty simple to put together. I'll be playing with img2img and trying out the IP-Adapter to see how it is the next few days. Thanks for confirming, good to know it's not just me. I hope they plan on fixing the artifacts though. Kind of a deal breaker for long term use.
ЕAnother model is good. What is known about CivitAi? Why did they freeze SD3 and are in no hurry to open it even after changing the license?
Apprently there are still concerns for enterprise license. Here is a good video that breaks it down th-cam.com/video/23tEjvCNZHA/w-d-xo.htmlsi=BGAmCeGL4jt02vEz
Thank you for your tutorial. How come my GPU is not used when it generates images?
If I have 11Gb VRAM my KOLORS-workflow will not work?
hi ..... can you tell me the locations of models after download.... cus I can't fiend it
The output looks better than Aura Flow (strictly going off your videos). I've not tried it as I'm not really a fan of ComfyUI.
@@cyberhard well keep in mind this is a fully trained model and Auraflow is still in beta and not fully trained. Prompt coherency is better with Auraflow I’ve found so far but again, different styled base models and both are fun to use. Will be very interesting these next few months
Yes finally!, although i think auraflow has a better promp adherence,i would say the photorrealism in kolors is better, now you could try hunyuan😝. I liked that You mention about celebrities, this new models seems to not be a big fan of them sadly
hahaha yeah that's another one I can do a video on. There is even Ella which looks interesting. The celebrities are a bummer but I did notice it knows Asian celebrities. hehehe. It's all good and fun for me, kind of a nice change from SDXL and while we wait for the updated SD3 medium.
@@MonzonMedia then My theory was right lol, of course it knows about asian ones haha, sadly i don't know any beside Lucy Liu xD, and after that ,You could try 900m pixart to SEE how it goes 👌🏻
Already recorded a video on Pixart Sigma and the 900M variant but I didn't like how it turned out. Going to do something else and re-record. Pixart is my least liked model out of everything I've tried lately but I still enjoy using it.
@@MonzonMedia ohh... That's sound like trouble, cant wait to see video👌🏻
I really want to try this model but the auto download won't work for some reason.. I've tried manually downloading but the workflow errors out and says it can't find the model.. I tried placing them in the suggested folders on the github page but no luck..
What are the errors you’re getting?
@@colinfransch6054 I uninstalled the node, let me try it again and I'll let you know
Strange, works for me every time. Keep in mind it can take a few minutes to download the models. Also, you can check your folders of the models are there, it's located in a different place than usual. For me my address is like this, yours should be similar. Note it should be under diffusers if you do the auto download node. C:\Users\ermin\comfyui\app\models\diffusers\Kolors\unet. In the same Kolors folder there is a folder for the text encoder, you should have a few models in there as well.
@@MonzonMedia I'll give it another shot
@@MonzonMedia I'm trying to paste in the error but YT thinks it's spam and is deleting my reply..
TY!
You’re welcome! ☺️
Day 1 of asking when will you make a video explaining Free-U ?
AuraFlow значительно уступает Kolors. Я тестировал обе. Kolors на текущий момент - одна из лучших генеративных моделей для SD. Но пока есть одна проблема - с ним не работает InstantID, но это, вероятно, скоро починят.
Well keep in mind Auraflow is only in beta at the moment and it is known that it requires further training. The prompt adherence is also a bit better than Kolors and close to SD3 coherency. That being said, I really like Kolors considering it's a base model BUT the artifacts is a major flaw in my opinion. I hope they release an updated model that corrects that.
@@MonzonMedia Согласен, что артефакты - это проблема, но будем надеяться на лучшее. Сейчас это спокойно решается hires. И да, претензий к бете нет. Ждем новых моделей, ведь в каждой можно найти какую-то прелесть :)
Very true! I'm having fun using these new models as you said each one has a unique style and to be honest I'm kind of bored with SDXL models as they are all very similar.
@@MonzonMedia Ну, это неудивительно, ведь все они, по сути, потомки одной модели :)
Do these new models only work with comfy? I'm not a fan of the spaghetti.
As far as I know only ComfyUI at the moment, there is an online demo where you can try it. huggingface.co/spaces/gokaygokay/Kolors. I haven't searched around for the latest posts yet but will see if there is any news on support for other platforms.
@@MonzonMedia Thanks for the response and your informative videos!
You're very welcome my friend! Appreciate the feedback and support!
Theses little "imperfection" seems like a canvas texture that you can see on real paintings. Maybe they trained it with ... paintings ?
I suppose it's possible however the artifacts you see are likely due to low quality images, synthetic images in the data set. I'm kind of surprised that they would release this way. I hope they plan on improving it.
@@MonzonMedia yeah, you are right ! Great video by the way, thanks for making them.
@@hleet you’re welcome! It’s fun to play with these new models 😬😊
Can this model also used in Forge ?
As far as I know not at the moment. It's similar architecture as SDXL so supporting it wouldn't be too difficult.
Actually MZ is the better version, you get access to more Samplers options,
The only problem I'm noticing with Kolors, is it doesn't like to do full body humans, maybe I'm prompting wrong
True, haven't gotten too far with testing other samplers (and nodes) as I just started playing around with Kolors these past few days. Will be fun to discover it's capabilities. As for full body, always helps to prompt things more specifically, like "standing on the side walk" or describing pants, shoes, ground etc. Full body isn't the best word to use and sometimes will give you a "full figured" person 😊
I often have more success with 'full body view'. Also interesting is 'full body view from side' or 'from above' etc. However, I have not tested this with Kolors yet. 🙂