Thank you for talking about our model! Glad you're a fan, and glad you're talking about Comfy! This is by far the preferred way from us. This, coming from an Auto1111 fanboy who has recently switched over...
@ischeka
Жыл бұрын
Do you have a nice way to browse Loras in Comfy, or do you just remember the names ? I tend to horde a lot of them and the external networks browse in A1111 is what is keeping me using it
@sprinteroptions9490
Жыл бұрын
I converted too :).. my wimpy laptop w a 3050 mobile 4gb will do a batch of 4 sdxl in 200 seconds. it's not bad bc you can queue up as many jobs as you want... just ctrl-m, boom.. tweak, boom.. then just review your prompts in the mgr later. by the time you looked at your first batch another's done. boom.. wonderful stuff
@notDreadful
Жыл бұрын
Amazing work as always! ComfyUI it just works so well, with A1111 it would take a long long time to generate a single image.
@CharlesOkwuagwu
Жыл бұрын
@@notDreadful❤
@Elwaves2925
Жыл бұрын
Nice video Olivio. FYI, if you don't want to drag and drop an image into the workspace, you can use the Load button.
@OlivioSarikas
Жыл бұрын
Thank you. Draging works just fine with images :)
@beatemero6718
Жыл бұрын
I figured out the drag and drop yesterday and was amazed that my picture, generated in automatic1111, had the workflow (with loras!) to recreate the exact picture. ComfyUI is absolutely amazing.
@davidw6422
Жыл бұрын
I rarely comment on anyone's videos, but this is excellent, i used SDXL in a1111 and it took 17mins for my RTX2070 to render and image, now it's under 12 secs. thank you for sharing your knowledge and taking the time to record easy to follow guides
@atanekoatan
Жыл бұрын
Hi. Do you render photorealistic humans? I am trying to get same results that I have with A1111 (which are amazing) in ComfyUi and it just can't render even close to what I used to have. I am really puzzled. I follow tutorials step by step - they work. I try my own prompts and settings from A1111 - and results are absolutely wrong.
@davidw6422
Жыл бұрын
I mainly do surreal, pop art, sci-fi fantasy, aliens etc. I don't concentrate on portraits, however have had some success with faces, I do feel a1111 is much better for faces, but I'm still getting used to CUI and feel it is getting
@benoitcorvol7482
Жыл бұрын
Did you find an answer ? i actually get almost the same but i kind of miss the interface of automatic1111 @@atanekoatan
@carnacthemagnificent2498
Жыл бұрын
I have a 3080ti (laptop version) with 16 GB VRAM and A1111 was choking on SDXL for some reason. It ran great ... until it didn't. Then reload, go for a while until the memory filled up, etc. So I tried Comfy UI. Running with the checkpoint and the refiner at same time and VRAM utilization is only around 35-40%. Amazing improvement for memory and speed when you consider it includes the refiner in each pass as well. Not sure I can stick with this UI and figure out things like extensions when we get them but for now I'm a believer.
@deathcometh61
Жыл бұрын
I'm on a 1660 ti laptop. A1111 was very limited. Comfyui has put out 1024 amazing photos. A1111 had issues at 768 half the time.
@Vyshada
Жыл бұрын
I had to do some manipulations with swap file because a1111 would crash attempting to load sdxl, but then it ran pretty good, though it will still devour all my ram. I installed ComfyUi extension for a111, and it have better loading, AND generating time. It's a little complicated for me, but already i can see just how much more control it can give me.
@Юрий-ч5м2ф
Жыл бұрын
Thank you for a very helpful and timely video! Unusual interface, but more flexible and advanced. Working with SDXL is much faster than in automatic 1111 and consumes much less RAM
@OlivioSarikas
Жыл бұрын
#### Links from the Video #### SDXL in A1111: kzitem.info/news/bejne/1qWLv42YfXmabKw ComfyUI: github.com/comfyanonymous/ComfyUI SDXL ComfyUI Example: comfyanonymous.github.io/ComfyUI_examples/ ComfyUI SDXL JSON: drive.google.com/file/d/1zZv9JG66CxMmR3sE7I-gAB2HkOFGYTZQ/view?usp=sharing SDXL Base Model + Offset Lora: huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main SDXL Refiner Model: huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/tree/main Google Colab ComfyUI: colab.research.google.com/github/comfyanonymous/ComfyUI/blob/master/notebooks/comfyui_colab.ipynb
@SpikeColgate
Жыл бұрын
Way faster than current Auto1111 implementation... SDXL is actually fun now! And automatically refining images etc just makes this way better overall...
@casualgamer3689
Жыл бұрын
Love your stuff, found out about the AI image synthesizers world just one month ago. I’m hooked and want to learn everything.
@TomiTom1234
Жыл бұрын
Awesome, ComfyUI is more stable and fast in my PC with 8gb vRAMs, and it does the refine on the fly. Thank you! All what I need now, how to train my own Model for SDXL beside LORAs.. Can't wait for your next videos on that.
@jonorgames6596
Жыл бұрын
May I ask, how much ram?
@TomiTom1234
Жыл бұрын
@@jonorgames6596 16 Ram
@Aesio92
Жыл бұрын
thank you :D
@paskmoe4969
Жыл бұрын
Great video. Cant wait to see some controlnet magic and lora/embedding training done on SDXL1.0.
@RegularInternetUsers
Жыл бұрын
That mouse callus is lookin' gnarly!
@MrPlasmo
Жыл бұрын
Wow thanks for another great tutorial! I’ve been using A1111 from the beginning and was not thinking about using ComfyUI until now to try out SDXL. What do people think? Stick with A1111 or is ComfyUI better for SDXL install?
@yessyspronck7483
Жыл бұрын
Autk user here.. im probably gonna switch over todag as there will be way less troubleshooting in comfy (i heard). Plus you will get a better understanding of the proces bc of the ui.
@mirek190
Жыл бұрын
and the best thing if you on windows you just have to download zip ready to work file.... no any additional extra downloads or installation
@raggadubsound
Жыл бұрын
Learning a lot with you, thanks for another video =)
@Steamrick
Жыл бұрын
Hey Olivio - check out changes in how prompts are weighed in SDXL. In SD1.5 giving a token a weight of 2 was asking for trouble - in SDXL you can go as far as a weight of 5 for some tokens and at least 3 for most.
@OlivioSarikas
Жыл бұрын
Thank you for lettings me know
@HolidayAtHome
Жыл бұрын
@@OlivioSarikas they said you actually should not go over 1, but instead go under 1 for stuff you care less about. So instead of "dog (with wings:2)" it should be "(dog.:0.5) with wings)"
@kinijustforfun5287
Жыл бұрын
How can i add face correction? That is really a problem? Hallo Olivio, bin ein großer Fan von Dir. Wie kann ich die Facekorrektur hinzufügen. Ich finde die Augen teilweise katastrophal, genauso wie die Hände leider auch noch nix sind...
@jeangranie3336
Жыл бұрын
Thanks again for this video. Is it possible to set weights for keywords in the prompt like in A1111 ? (or maybe another method to do the same ?)
@aidigitalmediaagency
Жыл бұрын
Wow, this is game changing. Thank you!!! ♥
@wpahp
Жыл бұрын
this json file doesn't have a lora node, is there an sdxl json that has lora?
@SFzip
Жыл бұрын
really appreciate the nice rundown 👍
@Van-Helssen
Жыл бұрын
Amazing! Thanks Olivio!
@MR3
Жыл бұрын
Thank you very much for this video with comfyUI its so much faster
@ysy69
Жыл бұрын
Your guide is really appreciated! Do you know if SDXL and fine tuned SD 1.5 models can be used in the same ComfyUI workflow? Like render an imagine using SDXL first and the apply img2img changes using SD 1.5 custom model ?
@grantmccotter8779
Жыл бұрын
Thanks for the great video! Do you know how to load in models that aren't a part of the original list? I tried pasting the url in the section with all the models and left out the #. It downloaded but it's not showing as an option in the load base checkpoint window. only the sd xl base and refiner is showing as an option
@freedom4sweden58
Жыл бұрын
Than you, now this works relly great whit my older GPU, a gamechanger :D
@michaelbayes802
Жыл бұрын
Hi Olivio, thanks as always for your great content. I've had a play with the google colab as described and it worked fine as you explained. The next step would be to try to install the ComfyUI Custom node manager somehow. Also is there a way of saving the models to the google drive so they don't need to be downloaded each time? I know it is pretty fast but if you have quite a few options and models that you need, seconds easily turn into minutes 🙂
@adriantang5811
Жыл бұрын
Many thanks for your great sharing!
@erdbeerbus
Жыл бұрын
great explained, is there an option to load a whole image sequence in comfy? using an IMG2IMG workflow ... thx in advance
@jcvijr
Жыл бұрын
Thank you! I will text this new method with sdxl
@iulianpartenie6260
Жыл бұрын
Thank you, Olivio! Where did you get pruned_no-ema?
@barni_7762
Жыл бұрын
I'm running it with free google colab and its actually pretty fast (~25s per image, a LOT faster than huggingface)
@burghardvonkarger5718
Жыл бұрын
How can I upscale with SDXL? I tried adding a second pair of KSamplers, but when I feed the output of the first refiner (after VAE Decode, Upscale and VAE Encode) into the second BASE KSampler, first thing I notice, there is no denoise setting. Second thing, If I run it I get a runtime error. So has anybody worked out already, how to upscale to 4k? Right now, the only way I know how is to use Ultimate SD Upscale with an 1.5 model, but that is, of course, not ideal.
@s.anonyme6855
Жыл бұрын
If you go in the link provided by Olivio for the Examples, you will see one dedicated to the Upscaling. Actually in ComfyUI upscaling is just 2 nodes (1 for the upscaling model, the other one for the operation). You can use that example to copy the 2 last nodes before the image and paste them in your own workflow. Or you can have a separate workflow for that. With ComfyUI, you can open several ComfyUI with the same CMD, and put different workflows in them (I do that). As long you don't use them in the same time it's fine. I generally have one for txt to Img, 1 for upscaling, 1 for inpainting.
@alinasama721
Жыл бұрын
Thank you! will you update a tutorial for how to train the SDXL model?
@changtc8873
Жыл бұрын
Hello Olivio, thank you for the updates. I'd also like to try out this wonderful feature. However, my graphic card is a GTX 1060 6GM. I'm wondering if this card is enough for this new feature?
@sherpya
Жыл бұрын
@Olivio where did you get pruned checkpoints?
@douchymcdouche169
Жыл бұрын
+1
@oliverhalsmanrosenberg6617
Жыл бұрын
is it just the user interface that is different?? (been using deforum on colab notebook) or are the results better?
@michealhsujejdud
Жыл бұрын
Hi dude your videos all very useful for me , i have one doubt about SDXL 1.0 AI images i can use commercial purpose
@Illusiongraphy
Жыл бұрын
How train models using SDXL1.0 with Automatic1111? I can't figure it out.. is there any way to train models with SDXL1.0?
@KeredaSmile
Жыл бұрын
Rember, after change path you have to "restart gui", co X and run gpu again.... a was use refresh and lost 2 hours to find a bug in my path...
@sigmata0
Жыл бұрын
Awesome. Thanks.
@blender_wiki
Жыл бұрын
I am surprised that 4080 is so slow, or maybe because you are on windows? Do you have pytorch for cuda 1.18? You use xformer?
@RSV9
Жыл бұрын
Are there also functions like image2image and inpainting here? Nice video, I'll think about it if I use colab. Thanks
@Steamrick
Жыл бұрын
Of course - both can be found in the ComfyUI sample workflows where Olivio got the SDXL workflow from.
@sancochodepapa
Жыл бұрын
Nice one , is there a way to do image to image prompting via comfy UI? That would be awesome
@OlivioSarikas
Жыл бұрын
yes there is. just use the nodes to build a image to image build
@yitz-e
Жыл бұрын
Awesome. But why is it soo slooow on an M2 Max mac with 32 gigs? Is there some setting or flag I need to toggle?
@greendsnow
Жыл бұрын
Hey! What if I dont want to use the GUI? I just want yo build the model and mass produce a list of prompts and save them with the prompt name on my drive... Do you have any idea how?
@bryan98pa
Жыл бұрын
First!! I love SDXL
@OlivioSarikas
Жыл бұрын
😍
@TheAries174
Жыл бұрын
Unfortunately i get repeated images.. Would lowering the denoising in the ksampler fix this problem?
@diecastkings
Жыл бұрын
you can also make ai video with this or only images ?
@lista_devideos_cursoscompl1338
Жыл бұрын
What a pain Robi alivee🎉
@JeffreyMerilo
Жыл бұрын
Great video as always! Can you please make one with img2img? Super thanks in advance! :)
@petEdit-h9l
8 ай бұрын
Pls how can i add the comfyui manager to the Google colab comfyui, pls help😢
@kelvinzhao4960
Жыл бұрын
Does anyone know if there is a way to train a Lora or Dreambooth for SDXL on collab?
@zhengsean3354
Жыл бұрын
How long does it take to render a image on Google Colab for SDXL?
@MrPlasmo
Жыл бұрын
When using ComfyUI, which folder do you download the SDXL Offset LORA into?
@travelpointone991
Жыл бұрын
Loras
@OnyxLee
Жыл бұрын
How to download checkpoints and lora from civitai to Google colab? Does this eat up your Google drive space?
@wtflolomg
Жыл бұрын
Ugh. My trackball's scroll wheel ... one click in either direction, the canvas is either microscopic or zooms beyond the canvas size
@silvermushroom-gamifyevery6430
Жыл бұрын
Heads up, Google colab blocks stable diffusion on the free plan, so you have to go pro or at least buy tokens.
@xlpcx
Жыл бұрын
Hi, how can I load some LORAS from cvitai for SDXL in the colab notebook?
@yazmo109
Жыл бұрын
why the pahin themodel the slash is facing the other direction? upscale_models: | models/ESRGAN models/RealESRGAN models/SwinIR while my path are usually c:\windows\user\ etc... they always have to make this not well explain what wouodbe the correct paths look like ?
@朱军军-n9l
Жыл бұрын
Why do I generate a lot of noise in the same engineering file, such as a poorly generated process diagram
@danielbernal839
Жыл бұрын
How do we can use the LoRA model from SDXL 1.0 in here?
@OlivioSarikas
Жыл бұрын
The don't work, and need to be updated to a SDXL version
@danielbernal839
Жыл бұрын
@@OlivioSarikas No, the LoRA that came with the Base SDXL 1.0 model, the one you downloaded in the previous video of Automatic 1111
@SANJIVRAI6693
Жыл бұрын
And here Im waiting to get this integrated in A1111 with 4gb vram capabilities
@mirek190
Жыл бұрын
lol ....A111 is not working properly with SDXL at all .. even in newest version creating pictures with SDXL is wrong.
@罗梦
Жыл бұрын
Has any of you found the comfyUI's zoom in and out is too sensitive with mouse scrolls. One tiny scroll can zoom out to max. Is there any solution to this?
@ICHRISTER1
Жыл бұрын
The results I get are weirdly RGB watermarked, like I can see lines and weird colos. Ive found online it is realted to the VAE, in automatic1111 I get none of the watermark (0.9 vae) Can I change the VAE in comfy?
@mirek190
Жыл бұрын
use updated SDXL models from 2 days ago -> like sd_xl_base_1.0_0.9vae and sd_xl_refiner_1.0_0.9vae
@TWJRPGGamming
9 ай бұрын
Can Controlnet run on Google Colab with SDXL?
@kenrock2
Жыл бұрын
Sometime I get stuck at "got prompt" like almost forever to generate every new prompt I inserted. It only gets fast if you regenerate the same prompt over and over
@KarolDecyk
Жыл бұрын
When i render this standard prompt with this bottle then it render really fast (i have geforce 3080 in my laptop) but as soon i change prompt then its take way way longer to render. Can someone explane me why? Can i fix it or am i doing sm wrong?
@santos.2020
Жыл бұрын
Hi thank you, is it possible to add infinite zoom to this?
@vallejomach6721
Жыл бұрын
Prompt executed in 531.94 seconds SD 1.5 it is then.
@shabadooshabadoo4918
Жыл бұрын
I have a 1080Ti and this is taking 115 seconds to render.. hmm I wonder if its just a card difference or something else.
@0BlitZKieG0
Жыл бұрын
I have no idea whats wrong with sdxl but it takes 30 min to complete image instead of models takin less than 1 min.
@mirek190
Жыл бұрын
what gpu?
@0BlitZKieG0
Жыл бұрын
@@mirek190 3080
@benoitcorvol7482
Жыл бұрын
Hello everyone, wondering why it took 25s for him with a V100 and myself 1min with an A100 ? ^^thank's for the video as usual it's very helpfull
@djevelyn
Жыл бұрын
I have a 8gb vram nvidia card, is it enough or will it struggle?
@gregorycoleman2149
Жыл бұрын
You can run xl on 8gb
@jaco001
Жыл бұрын
Olivio - do not blurry password or other crucial info. It is easy to reverse it.
@Gavanater7
Жыл бұрын
Do you think it will get faster with automatic1111?
@OlivioSarikas
Жыл бұрын
yes, for sure
@Rimbo28
Жыл бұрын
How can I develop a Stable Diffusion checkpoint from 1.5 to SDXL 1.0 ?
@jibcot8541
Жыл бұрын
What you have said, makes no sense.
@muledeer1766
Жыл бұрын
Where did you get pruned SDXL 1.0 ?
@fossil98
Жыл бұрын
It is pruned, its just still big.
@LpLelin
Жыл бұрын
Can you make Google colab Kohya ss tutorial
@RSV9
Жыл бұрын
It runs on an AMD laptop, Legion 5 and a 3050 Ti graphics card with 4 GB of memory. The times are much longer but it can be executed. Using run_nvidia_gpu.bat, I use sd_xl_base_1.0.safetensors and sd_xl_refiner_1.0.safetensors which differs from the ones used in the video which are sdxl_base_pruned_no-ema.safetensors and sdxl_refiner_pruned_no-ema.safetensors and I don't know why . In the example of the bottle with the sky in the background, it takes 1:18 minutes and for the other example it takes 1:15 minutes.
@AgustinCaniglia1992
Жыл бұрын
I don't understand why this workflow doesn't have denoising on the refiner step.
@lukeovermind
Жыл бұрын
I am might be wrong but my understanding is that, you splitting the denoising between the two samplers. Let's say SDXL base at 20 steps and Refiner the last 5 steps. The Base Sampler leaves some noise for the refiner (5 steps) to do it's thing. This is all happening in the latent space. Where with Image2Image a pixel image is taken as an input and then a certain level of noise is added according to denoising strength
@lukeovermind
Жыл бұрын
Comfy is cool, the sampler has a setting where you select to leave some noise
@aceathor
Жыл бұрын
Invoke Ai do that, but I dont know if it do it with SDXL
@krystiankrysti1396
Жыл бұрын
Comfy is slower than AUTO!11 by half, it does not use xformers properly or at all , i get 1.7 secs in auto11 but get 3.50 in comfy , same setings
@TR-707
Жыл бұрын
Wonder if this it for A1111
@PradeepSuranga-h4d
Жыл бұрын
why always need to download model ?
@MADVILLAIN669
Жыл бұрын
it takes me like 8 minutes to renders a single image in A111, I hope they'll fix this eventually
@mirek190
Жыл бұрын
lol ....A111 is not working properly with SDXL at all .. even in newest version creating pictures with SDXL is wrong
@generalawareness101
Жыл бұрын
Where did these come from? sdxl_base_pruned_no-ema.safetensors and sdxl_refiner_pruned_no-ema.safetensors? I do not find those on the HF page you linked.
@TysonChenys
Жыл бұрын
Can m1 Mac mini run this?
@Osama-xs8cl
Жыл бұрын
Can I do img2img or inpaint in XL ?
@mirek190
Жыл бұрын
yes
@Feelix420
Жыл бұрын
Config file download button is not working
@stas_lu
Жыл бұрын
on my 1080 video card, 1 picture is generated for 1 minute. it's a very long time. (
@mirek190
Жыл бұрын
You have 10 years old card ...what you expect?
@stas_lu
Жыл бұрын
@@mirek190 on 3060 video cards the same speed
@ruuuuudooooolph
Жыл бұрын
be thankful, mine on comfyui takes 16mins
@toygunsonly8093
Жыл бұрын
Read this way "Comfy UI" Not this "comfyu I"
@SteveM45
Жыл бұрын
hi, with my RTX 3070, 8GB it need 18 s
@Bhiku_mhatre
Жыл бұрын
Please get roop for sdxl 1.0
@Smashachu
Жыл бұрын
Alright off topic, but i can tell you use a mousepad with a wrist rest, because the side of your wrist on your mouse hand is inflamed.
@ragnarlothbrok367
Жыл бұрын
i dont like comfyui, it is not comfy at all
@gregorycoleman2149
Жыл бұрын
Comfyui is how software developers make art lol. I prefer invoke ai.
@mirek190
Жыл бұрын
I like it .
@gregorycoleman2149
Жыл бұрын
The fastest way is to use a 4090
@a_vin7644
Жыл бұрын
I'm so lazy. I want be artist and not engineer. I need application with only few knobs and regulators. I want think about image and not about meaning of 1000 knobs. For example there is great application Corel Painter with so many possibilities for brushes so instead of create I waste time on creation brushes. But every one can choose his way. Enjoy be creative
@manitou__
Жыл бұрын
coupy,lol
@ameet21
Жыл бұрын
how you created sdxl_base_pruned_no-ema.safetensors and sdxl_refiner_pruned_no-ema.safetensors? Please make video and please make video on where we can merge our model with SDXL
@OlivioSarikas
Жыл бұрын
These models are created by stability, the company behind Stable Diffusion
@cantoversial
Жыл бұрын
@@OlivioSarikas thanks for the tutorial. tried running comfy UI following your instructions in Google collab with your JSON but getting the following error: Prompt outputs failed validation CheckpointLoaderSimple: - Value not in list: ckpt_name: 'sdxl_base_pruned_no-ema.safetensors' not in ['wd-illusion-fp16.safetensors'] CheckpointLoaderSimple: - Value not in list: ckpt_name: 'sdxl_refiner_pruned_no-ema.safetensors' not in ['wd-illusion-fp16.safetensors'] did I miss some steps on installing those? appreciate your help with this 🙏
Пікірлер: 161