As a novice in all this, you would think with ALL of these downloads and distribution requirements that 'they' would have a simple batch file to download all of the packages and give them the right names. I don't understand how there can so much work put into development, yet for someone that knows what they are doing would take them 10 mins to write a batch script. Just think everyone that wants to use this IPadapter Version 2, every single person has to follow this monotonous process (and this is the EASY install guide!)
You say that, but it is still work. You have to write different script for different OS. You have to adjust it each time you make changes on your project. You have to respond to issues people have with it. You have to test the script... Timesinks add up, this IPAdapter is just from one person. but you know what, GPT4 and Claude Opus exist. So you could do it too with trial and error and share it with the community.
After a month of trying and reading the forums, you solved my problem. I don't know why they don't explain clearly and clearly like you on the main page. Thank you for providing us with your information. 👊👊
It's because SD people bought comfy and made it official SD app. Still, a1111 gets more stuff from community as it's more popular and bilion times easier to use.
Basically the midjourney consistent face feature, but it's: 1. Free. 2. Open-Source. 3. Without restrictions. 4. Arguably looks better. 5. More freedom with the models/styles you can choose from. That's great!!
well, the model is from a company with a lot of restrictions and pretty strict licencing rules as far as i know. So not that. But for private purposes it is very powerful and fun
It is not hard to fix existing workflows to use the V2 IPAdapter. Once you upgrade ComfyUI with the new nodes, you just need to load the older workflow, delete the IPAdapter nodes and replace them with the new nodes, then refresh ComfyUI and reload the browser window and the node should run as usual.
He hasn't covered webui in months. There are plenty of IPAdapter videos for Automatic 1111 if you do a search. I don't get why everyone keeps complaining about ComfyUI, it would be like going on a Houdini tutorial and complaining that they're not using Blender when there are hundreds of videos on it already.
Because comfy is over hyped overcomplicated garbage. Only con of comfy over a1111 is that it turns on faster. A1111 is billion times easier to use and gives exactly same results. Comfy was better when it was new as it was faster but now it's not the case.
Beware that the automatic update of the manager sometimes doesn't work and you may need to upgrade manually. This is NOT an addition but a complete replacement of all files
Does the simple build example you showed first also require Insightface (with its copyright limitations) or does this only apply to using the Face ID tool?
Great video as always, Olivio. Even though I catch AI news here and there, I'm completely lost these days. The last thing I was familiar with was Stable Diffusion 1.5 and A1111. But now I come across this video mentioning ComfyUI and other strange things I don't understand - where did those come from and when? As a suggestion to help newcomers break into learning these technologies, it would be helpful to have a summary video every 6 months or so covering the current general use tools. Everything is moving so quickly that I no longer know where we're at. Thanks. Liked and shared.
You're pretty much the only AI youtuber who is consistently staying on top of everything at this point. Thank you for keeping us up with the new developments🙏👏👍
Comfy is just SO HARD to use. It feels more like coding with having to permaneently connect and deconnect multiple nodes to just upscale and use a fkin controlnet… but it seems to be the most optimized one though. I’ve tried a1111 swarm and forge and only comfyui is able to run lightning models on my poor 2060 the others just crash….
Many might ask what IPAdapter actually does and unfortunately this was not addressed in this video. The IPAdapter are very powerful models for image-to-image conditioning. The subject or even just the style of the reference image can be applied to a generation. Think of it as a 1-image lora.
@@OlivioSarikas It doesn't work if you show the results at the end of the video - most people won't watch the entire video when they aren't sure what it is about in the first 5-10 seconds.
Note: if yours was working and now isn't working its likely because you deleted, moved or changed where the original file was you downloaded in this process olivio shows. Most of these issues can just be solved by leaving the file where it was originally downloaded lol
No, it has a couple new settings that affect how IPA is applied. I've gotten better results with the old full-face models on stubborn anime models that didn't work before, I haven't tested other models yet but it is a big improvement. That's on the advanced node.
Hi, I tried so many times but there is no way to see Ip Adapter plus node... I really don't know what else to do... The strange thing is that from terminal it said the node load was successful....
pain of installation is your last layer of protection. Because when its gone, everyone will have AI in their phones. And noone gona need you anymore :) So enjoy current days while you can !
@@stranostrani9212 Many reasons... Python is not meant to be used globally unless you know what you are doing. By making PATH collide like this you will introduce hard to debug issues. Most tools just take the first matching python it can find in PATH, wheels might install into the global path but using this local version etc etc... Just type "why use a virtualenv" in google if you want to know more
I did everything you said, put everything in the correct folders, installed everything, got no errors and still can't see the IP nodes in comfy and I don't know why 😔
@@Zibalaka1 The new IPadaptor has been rewritten and I think as the main dev has stated. There will be many breaking changes. So it will be better to wait until a stable release is present. If you still want to go ahead, I suggest using the ComfyUI Manager to install missing dependancies and read the error messages to see what libraries are missing. Then you can google them and install accordingly.
Extremely Inconvenient. I will refer to this tutorial when I innocently update my ComfyUI and discover that all of my workflows are broken 😃. May solve a lot of frustration in future! - (Thank you in advance) Which real usage issues was this total rewrite meant to actually solve? - Had no issues with IP adapter usage whatsoever. - Not complaining ofc.. Just that, as you said 'it's a pain'.
1. speed improvements 2. fixed all directml issues 3. easier use bc you won’t have to remember and connect the right clipvision and ip adapter combos and weither it’s a plus and or v2 model. 4. being able to use non-square inputs. Clipvision only understands squares, the new ip-adapter tiled automatically crops the image and generates according masks. That’s about it for problems, everything else is extra features, like optional negative image in, being able to select the mathematical approach of how multiple images get mushed together and weight types (slowly fading in/out the ip adapters influence over time thereby giving more control over it than just start and end time). Also you can now use masks for the embed loader, not an attention mask , but a regular mask for what part of your reference images should be considered by the ipadapter. It’s especially strong with rembg, be it objects, animals, people, clothes, removing the background removes its unwanted influence on the output. So overall a Game changing Update if you wanna explore the advanced new features, if you just need the basics tho it gotten kinda more complicated if used to the old way but I can imagine it being more beginner friendly as people won’t have to hassle with all the different possible model and setting combinations. On the other hand, there’s even more possible settings now if you wanna fiddle with different weight types and embedding combinations
You can go into the manger and update them one by one as you did before. Though if you are worried about yourworkflows, the iP adapter itself is different enough that your exising ip adapter workflows own't work after updating it
Hi Olivio, do Ksampler node and Ksampler Efficient function differently? I made the same initial workflow / settings using Ksampler efficient, but not getting the same output as you have shown
@@OlivioSarikas Was just studying this more, do you think there may also be an issue with Efficient Loader node? I don’t think the prompt gets transferred to the Ksampler the same way as it would through clip text encode > K sampler +ve conditioning
Side note, I think you're using the wrong version of the IP Adapter (you're using the IP Adapter Plus node which Mateo deprecated in the latest update, it's only IP Adapter Advanced node now)
I'm not using the IP Adapter Plus node, the Nodes Pack is called "IPAdapter Plus" - As far as i can see i use the same nodes that Matteo uses in this Video and he also links to the IPAdapter Plus node Pack: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-_JzDcgKgghY.html
@@OlivioSarikas OK I may be wrong, but if you use ComfyUI Manager and update the IP Adapater node pack, see if you can add a regular "IP Adapter node" to your workflow. I only see "IP Adapter Advanced" but again apologies if I'm mistaken
@@GrocksterRox I had the problem with not seeing the plain IPAdapter node, but Matteo explained that you need to delete the "IPAdapter-ComfUI" folder under custom_nodes (not "ComfyUI_IPAdapter_plus", of course!). This was by a previous developer and is deprecated, with relevant code moved to ComfyUI_IPAdapter_plus. After deleting this and restarting ComyUI I could find the regular node.
The "Weight_type" in "IPAdapterAdvanced" does not have "style transfer(SDXL)" and "compostion(SDXL)".Other functions are working fine, but there is no "style transfer(SDXL)" and "compostion(SDXL)", I would like to have a separate reference for styles, but I can't get it to work. How to solve this problem?
Epic walkthrough! You made navigating the IP adapter version two feel like spreading butter on warm toast - smooth and satisfying. Eager to dive into the unified loader's buttery goodness. But really, who's truly prepared for the sheer, unadulterated power of Olivio Butter? It's like the secret boss level of kitchen spreads! 😂🧈
Easier but less capable ;) That’s the tradeoff with auto/forge. IPA has been working in a11 for a long time, but the new update is comfy exclusive yet and I don’t see Matteo spending his little spare time to go horrendous efforts in making it work with auto, if anything it’ll be on forge first as that Plattform is much better at allowing devs to easily implement their new ideas. Auto is lacking the different weighting’s, the negative image input, the different choices of how embeds get calculated and the flexibility of comfy. There’s a new inofficial ip-adapter for composition, its not implemented in any gui, so I added like 4 Lines of code to the ip-adapter-plus comfy custom nodes and now I can use it with the unified loader and it automatically selects the right Clipvision model. If you don’t get the benefits of comfy you probably won’t benefit from it as its admittedly more difficult to start with, but it opens up endless possibilities if you are open to learning a bit of logic and like following a linear path of execution
So, if the image I'm using has both a style AND a face I want to reproduce, I should just run the two loaders into each other and then run the second one into the Unified Loader?
why everything has to be so stupidly dificult just tu run some nodes. i'm a photographer and i really need to use these tools for my job and clients but i don't know how to code. hope some day someone will come up with an easy solution to use these tools...
Is faceid better than giving an already cut face to ipadapter ? I have an entire workflow to manage gender detection in an image with 2 faces to process in 2 different reactors (reactor gender detection is bad) and after that I am using ipadapters to make the faces less cartoonesque.using the faces segs and masks I got during gender detection I never redetect faces in a large image so it is faster. Is the quality the same ?
yes, i have and i made a 30 minute video for my patreon supporteres with different cloths styles and also with background combinations www.patreon.com/posts/ipadapter-v2-101309573
I am already running IPAdapter workflows... if i update, will i have to download the models again ? or will the update automatically reference them into the new nodes ?
made an update through the Manager (Update All), got an error "When loading the graph, the following node types were not found: IPAdapterAp" Does not help "Install Missing Custom Nodes"
You have to delete the old node from the workflow, replace it with the appropriate new node of the V2 IPAdapter, then refresh (and or Restart) ComfyUI, and reload the browser window once it restarts and the node should run as usual.
not working for me .... have the same settings as you have... only differnece is that I am using a lightning model... but all I get is a deformed winter jacket .... no woman.... or any subject
every time you make a video, I never see you hit generate as if you want people to buy in to your patreon... This is the second tutorial of yours that has come across my recommended feed that didn't work... I don't know how you got this subscriber count but I'm done with the video baits you do. I am not interested in this channel going forward!