InstantID for Automatic 1111
Вставка
- Опубліковано 28 вер 2024
- Instant-ID for A1111 is out. Here is how to install everything. I show you the settings I use and the Results I get with InstantID in this Tutorial
#### Links from my Video ####
/ oliviosarikas
civitai.com/mo...
github.com/Mik...
#### Join and Support me ####
Buy me a Coffee: www.buymeacoff...
Join my Facebook Group: / theairevolution
Joint my Discord Group: / discord
AI Newsletter: oliviotutorial...
Support me on Patreon: / sarikas
Finaly we can have some of our sanity back with A1111. Comfy ui was driving us all nuts.
No it wasn't. ConfyUI is lightyears ahead of A1111 ;)
ComfyUI grows on you :)
@@OlivioSarikas Yes it was. I even unsubbed you after watching your videos for a year, because it was ComfyUI everything everywhere all the time. And I AM an IT guy, I've been using Linux for 30 years, and I can code in multiple programming languages. I just don't want to have to deal with that Spaghetti nonsense. Tell you what: Put a GUI over Comfy, one that hides the Spaghetti and exposes only in and outputs, so that nobody has to deal with it who doesn't want to....
Re-subbed, btw, because I do like your videos after all 🙂
@@OlivioSarikas it still can't do some basic stuff I need, which A1111 can. So ..
@@OlivioSarikas Was going to ask, these same Controlnets and similar configuration should work in ComfyUI correct? I've been watching a lot of your ComfyUI videos and I have been loving them! I haven't tried A1111 yet but considering giving it a little experiment
I can't get into comfyui rn, so this is very welcome. I'll get a Linux setup first for comfyui
I think the 2nd controlnet doesn't have to be the same image, can put any image in there that is sort of used as a pose, basically takes the position of the face. Something that also helped the speed was going into A1111 settings and set the controlnet model cache size to 2 instead of 1. Sadly I guess this doesn't work with normal hiresfix but it seems to work with Kohya hiresfix although I don't have much experience with Kohya's version.
Correct, the second controlnet image can be any image.
Came here to say this
Thanks for the extremely useful videos as always. I tried to follow the discussion on the Github repo for the A1111 extension and failed to get it working correctly, but your video made it a breeze.
@OlivioSarikas, that's great! The current ComfyUI nodes for Instant ID are very restrictive, because you can't connect loras or others ControlNets to the pipeline. I wonder why the that A1111 extension is more flexible and the current comfy UI nodes are not.
Hi, I think this is because comfyui's node implementation now is based on diffusers backend. So every node based on comfy backend won't work.
Love your videos. I've only just got into all of this and find 1111 much more accessible as a beginner. ComfyUI is the end goal, but I'm glad you're still covering 1111 tips, too!
I had the same thought but found ComfyUI turned out to be easy. Especially after following Olivio's ComfyUI tutorials on this channel. I ended up regretting the time I spent on Atomic1111. Don't wait too long before seriously trying ComfyUI.
Thank you for always explaining in an easy-to-understand manner!!!!
i was waiting for thisss....Thank you !!!
Is this an alternative to making a LORA and does it include the body?
Wake me up when there is a version for SD 1.5^^
And when it’s comercial. Though I will say that SDXL’s latest community-trained models are a lot better than they were at first.
exactly, I don't have a grand for a gpu. I have kids and a wife that are my priorty. This is great as a hobby but feeding and clothing my kids rent bills aren't cheap.
its trash, dont waste ur time.
There already is, called FaceID. V2 works as good as 1.5 can master.
Hi Olivio :-) Finally for A1111 - thanx for the Video. But pls help me.... is this nice tool for restyle only or is it also possible to use it as a reactorlike tool? which means - more complex compositions beyond a portrait?
Always Love your videos as they are informative and fun to watch. BTW.. why do you have 2 ControlNets? ControlNet and sd-webui-controlnet?
When I do this in Auto1111, nothing happens with the controlnet. It just generates an image without anything to do with the Control0 and 1 images.
Hi, I doiwnlowaded the two files and renamed as required... but Auto1111 doesnt recognize the Instant... file. It does recognize the other one (adapter) the name I already checked is correct, it is a safetensors file, what do you think could be the problem?
Thank you!!
Hey Olivio, great video as always, but one question - when I download those two models and rename them, what extensions should they have - .ckpt, .saetensors, or no extension? Just cannot get this to work and want to eliminate any problem I can.
the same they already have. you just rename the file, not the file type
s @OlivioSarikas when I download the file is a .bin file is this correct? All others appear to be safetensor
Does fooocus have this feature ?
It works but the final image is totally burn out... any ideas?
they recommend setting the cfg lower, 4-6
I was playing around with this a bit more. There's a realvis 30 turbo model that works really nice with this. If you set cfg to 1, and steps to 5, it works perfectly
@@generichuman_ Thank you. I will try.
Once I get more RAM I'll be rocking so much with the SDXL
It works for me with RTX 3060 6Gb vram. Just remember to use --lowvram when you start A1111 and check Low VRAM in both controlnets
@@brunodegiovanni6428 Oh dang. Thanks!
waiting for SD1,5 version
How to get rid of the fact that instant id constantly changes the direction of the eyes, the eyes always look either down or sideways. In promt I always specify looking at the camera, but it doesn't react
not using automatic1111 for a loooong period of time and I DO not REGRET it !!!
One day I'll be brave enough!
Olivio, many thanks! I am now experimenting and comparing the results with those generated by the gradio demo on huggingface. The "IdentityNet strength" seems very much related to the ControlNet0 weight. Do you happen to know what is related to the "Image adapter strength"? It does not seem corresponding to ControlNet1 weight, as on the demo I am not even using the optional pose image!
thought you forgot about us a1111 users!
So if i made a character and want its face i just use this and have the same face?
Yes
May I use this tools to edit the photograph instead of changing the image look ?? For example change the cloth or the haristyle ?
you can probably do that with inpainting alone
Hi Oli !!! Can you make a video explaining how to install xformers in the recent Stable Diffusion A1111, the videos i found are old, and find people saying they got errors installing xformers with the old methods in recent updates. Thanks!
it should come default with new versions of auto1111 just add the lauch option argument to the bat file. i use:
set COMMANDLINE_ARGS= --medvram --no-half-vae --xformers
Do u recommend replacing turbo model u used with the "sd_xl_base_1.0"? Does it have same data like the base and is also 'turbo'?
no, a turbo model is made to work with low step counts. the base model isn't even trained on interesting images. you get much better results with community models
@@OlivioSarikas thank you so much.
Exciting Matt! Thank you 🔥🔥🙏🔥🔥
Thank you for this video, this is super nice. I tried doing this, everything works but it takes 3min to make one image at 512x512 resolution. Is this normal? I am using TurboDiffusion model.
Is there a workflow in confyui?
Huh, I was placing the image I wanted to place the face on in my second controlnet - had to experiment a bunch but was able to get the result I wanted - my face in controlnet 0 merged with a diffused version of the image in controlnet 1. Was this the incorrect use?
the face you want should be in first controlnet and 2nd controlnet acts a bit like a pose, basically taking the position of the face in the 2nd image, so the 2nd image doesn't have to be the same image as first image. This isn't like quite like faceswap or reactor where you put the image in that you want to replace the face with. It allows more style/creativity so more based off the image created from the prompt/model and then it's able to take the face from 1st image to work it into that result.
i need those wildcards!
Thanks!
Just fyi to everyone don't change the file extensions! Just the names.
It's very nice but why just for XL models?!
SD1.5 is more popular and usable right now.
Hope for SD1.5 soon
Can it work on furry?
How do you get the instant id tab in A1111 control net when update does not show it and how do you get the preprocessor for instant id to show up? still not working :/
It seems like you still have an update to do. You need ControlNet v1.1.439 and the tab and preprocessors will show. You only need to download the models. If that doesn't work maybe check what version of A1111 you're on. With v1.6.0 it works. 🙂
@@Elwaves2925 the check for updates button was not working, i had to delete controlnet folder and install it again :D
@@farfromerease Happy you got it working. 🙂
This just takes forever even with turbo, and the likeness isn't that great. With ipadapter faceid portrait, I get much better likeness in less than 10 seconds.
fair warning this is using the inside face technology so you can't use this in a commercial way ...
what do you mean by this? who can stop me if i do use face recreated using this in a commercial way?
the law? their lawyers? that's what licensing is for
@@OlivioSarikas I mean that has to be country specific thing....and how will determine the image is ai generated..granted that if we use face of a popular celebrity..
Or a political or an actor that will definitely create problems..but what if I use my own photo ..will that also create a problem .....?
I heard that the instantid uses like 20 GBs of vram
Is it normal for an RTX 3070 with 8GB VRAM to generate an image in about 4 minutes?
No, I'm using a 1080 Ti 11GB VRAM and a i7-8700K and it always take 1.5 minute.
Did you try to launch A1111 with these parameters ?
webui.bat --xformers --no-half-vae --medvram-sdxl
Also, you need to set A1111 Settings named "Model cache size" to 2, save and restart.
2080 8GB VRAM. 6 - 12 minutes.
I do not understand the benefits of this now compared to IMAGE2IMAGE with the A1111 controlnet solution?
If i add a picture to image2image and use style lora or add a prompt style it also looks like the character in that style?
I thought with instandID i can use a picture and use the character then consistent in all angles front side and it looks always identical so i can create a story then like
a comic and the character always looks 100% identical?
Face,clothing and so on.
Can i i use this Instant ID Character to work with that other than adding a style?
I also want to create face textures with AI to create unity models and for that i need
1 pic close shot front (exact matching angle ) as the video shows from minute 3:21 on
1pic close shot from side left ( exact matching angle )
1 pic close shot from side right ( exact matching angle )
In high resolution which is impossible using a LORA which creates multi angles in "one pic".
The output cannot be used if you cut these out and add it to a mesh due to its low res.
I need two, better three single created pictures of same "identical" face at least front and side angle in best possible resolution.
So can Instant ID do this now finally in AI World to keep the person as it is?
Auto1111 is best. Comfy ui is made by 16 yr old who had nightmare 😂 it confuse everyone
InstantID Portable more fast
It takes forever to generate an image lol
Meh
People are still using A1111?!?!
using only the first controlnet works the second one is just to give exact pose as the image. you can also use a diffrent pose image in the second controlnet and a face in the first, you can also use open pose instead
Great tip!
thx for this i was going to ask if you were limited to the pose on the pic. cheers!
YESSSSSS!!! Finally another A1111 video. We don't do COMFY-roun-heya!
Comfy gang gang.
thank you for doing automatic 1111 stuff again
Nah we need to go back to Comfy.
no@@J3R3MI6
Does this work for forge? I do not have the instant id face keypoints preprocessor and I did your methods. Also another yourtuber who is trying to do the exact same thing as you recommends using instant id embedding..very confusing why to preprocessors for same thing. Which one is better?
so this won't work on models based on SD 1.5 then?
Thank you so much for doing these in Automatic as well. You are always awesome mate🧡
Don’t forget to set Controlnet cache in A1111 settings to “2” for cache each controlnet model that we are using. It will make rendering so much faster after first run. 😊👍
Good call!
how do i do that?
yes, how?
But what side effects will have this, need more ram, vram or?
@@ZeroCool22 of course you will need more VRAM! ;)
Should teh files that i am rename be still .bin files, or do i change them to .pty ?
i just updated controlnet and now i have a error "attributeerror 'nonetype' object has no attribute 'values'" please someone help
Great news -InstantID in A1111 - but how to use openpose with that? - right now I get only 1:1 portraits like the input image...
I'm having the exact same issue
InstantID for Automatic 1111, it didn't work, it doesn't generate my face, it turned everything on and updated, who can help?
Having the same issue
you seem to have the model ip-adapter_instant_id_sdxl and control_instant-id .....i have both the id preprocessors but neither of the models came with controlnet update that you showed, can you tell me where to dl the models and place them?
He shows you where to get them in the video and what to rename them as (very important). The link is also in the description (third one down). You put both of them in your extensions/sd-webui-controlnet/models folder.
@@Elwaves2925 tx alot mate i usually follow these installation instructions as their explained on my second monitor, pausing to make sure i do everything right, but he shows you how to choose them and then explains where to dl them afterwards in this video which had me and a couple others confused lol cheers..
@@thebonuslvl No problem and you're welcome. I do the same with the pausing but experience has taught me to always watch the full video once, then go back. It can save a lot of hassle. 😁
You also can do this with FaceID through IP-Adapter (working also for SD1.5)
can't post link because message will be deleted
@@HappyBowser777 ip-adapter-faceid-plus_sd15.bin and also need lora: ip-adapter-faceid-plus_sd15_lora
@@HappyBowser777 ip-adapter-faceid-plus_sd15 and also need lora: ip-adapter-faceid-plus_sd15_lora
@@HappyBowser777 ip-adapter-faceid-plus_sd15 with lora ip-adapter-faceid-plus_sd15_lora
@@HappyBowser777 FaceID Plus with FaceID Plus LoRA
i can't seem to get this working on forge. do you have a video for that?
Same, it won't work for Forge
could you please tell me what gpu you are using ?
I don't know if I'm doing something wrong, but for some reason, I'm not liking this InstantID compared to FaceID sdxl + IPadapter Plus Face sdxl (0.2 weight). Not only it is more resource heavy, but also achieves less likeness to the person in the photo (it works somewhat, but not as well as I would like). I really want this InstantID to work for me, but for now, I'll have to stick to other one.
Hi everyone, I would need your help to understand my mistake, after downloading and changing the 2 names of the models... I don't have instant_id_face embedding but I have the choice of instant face (instant id) Can you help me! How do I fix this? A HUGE thank you! :D
DId you find a solution? Have same problem
No i don’t :(
Bit confused with how both the ending control steps work
automatic1111
Excelent video ⭐❤
There's a reason why you didn't try a further away subject, isn't it?
Is anybody else experiencing a pretty hefty slowdown with this? I can generate an image with a turbo model in a few seconds, but with this CN it seems to take forever. (Around 5 mins each image) Is this normal?
Love it, thanks so much! It's insanely slow on my end. 60+ seconds / iteration, regardless of SDXL Turbo or not. I am using an older RTX 2080 Ti 11GB and am fine with any other types of image generations with or without ControlNet. Is Instant-ID that hungry?
Yeah, it's slow and memory hungry for pretty much everyone. It's still new tech so hopefully future updates can improve that speed. Changing to turbo did nothing for me except produce much worse results.
yes very very slow....
OK, thanks guys for confirming! Thought there's something wrong with my implementation.
It's very helpful for me thanks for the amazing content ❤❤❤
Thank you to explain to me because I had no clue Now I try to do it
Hey, need to know why all SDXL based Checkpoints takes too much time as compared to like picXreal and other checkpoints.
Why can't you use this for commercial use when you can use it offline? How would somebody be able to tell what you used to create a specific image? (Just asking for interest reasons - I'm not using any AI tools, and I'm not planning to use any AI image generator in the future, but I'm genuinely curious about this). Thank you!
I tried this, but a lot of the generated images come out blurry or looking very low res. Did you encounter these issues as well?
Try "ControlNet is more important" for both ControlNets
Try "ControlNet is more important" for both ControlNets
I have a similar issue.
can you provide with feedback I follow the process to istall instan ID but in the gui do not show the instantID radio selection in controlnet
I have better results with ReActor :D , Love your videos Olivio!
Can you share the prompts you used for the various samples?
is it better than ip_adapter_plus_face_v2?
it's not doing anything for me, i wonder if the model or something is corrupted, since i did have some stutters on the download... otherwise i dunno. The process is straightfoward, just another model for controlnet, only with two tabs... i don't understand why its not working
did you fix it ?. its not working for me either.
What about Inpainting? Can it work with a masked image?
It doesn't seem to be ready for that, right?
slow....
Good, very good video
Don't normally comment on your vids but always watch. Thank you for your great, quick, to the point videos. You are one of my go to whenever I am looking to delve into something new.
can it be combined with open pose?
This looks so amazing! Just wish not so complex to set all up
Thanks for youre good videos!
Thanks, wonderful.
👋
I can't get the ComfyUI version to install correctly. Keeps saying Import Failed (w/ Manager and GitHub Pull)
I'm sure, if a little more effort was applied, that the whole process could be even more long winded; I think I'll wait for grown-up program.....
Olá Olivio tudo bem amigo? Se me permite a sugestão, faça um vídeo falando desse modelo, o Turbo Diffusion XL Turbo, achei ele fantástico, faz um vídeo de com dicas para fotos, animes,vetores,logo e etc.. para o AUTOMATIC1111, abraços e feliz 2024 atrasado kkkk
is it work at Comfy UI?
This will be very slow by default! You need to go to settings > uncategorized > ControlNet and set Model cache size to 2 or greater. Otherwise, every image generation will swap a controlnet out of memory. Now you have normal speed.
thank you. i will try that