I think that the best "Blender Move" we as the AI community can make is to get our own fork of 1.5 or XL trained from the ground up with complete transparency as to where the training data came from. Not to appease detractors, but to legally cover users bases moving forward. Blender didnt start out free, it was Liberated.
Well said. I agree with you in this. Profissionalism is paramount here. The community has the professionals already, but maybe we yet lack proper motivation for doing such a serious endevor. Perhaps with Stability getting more and more closed we could trigger the right professionals for making this fork really compliant with other artists right's.
@@tiagotiagot it will be tough to verify content is indeed open source, so we should instead tokenize the base art and pay super fractional royalties and it remains that subsequent ai will be allowed to access via attribute based permissions, and if the viewer is willing to pay for entrance to a digital gallery and a fee goes to artists. Sames goes for use of base weights, except it's direct. No Spotify taking advantage of artists crap. Artists shouldn't even be left in the dark on pay, no patented idea should be used and not pay royalties imo. It's markets of scale. If all open-source images can be verified in the pre training it may be less complex than with royalties, but the community-source has advantages of higher quality input data most likely... eitherway the value of the art must be curated or all be left equal and see if the model figures enough out... but not just anything goes as far as data quality. It must be tagged and perhaps ranked for quality markers in scores on an index of various styles. How renaissance is it, how pop culture? Etc.
Open source is very important. Comfy Org is the only thing going on right now in AI that is really staying an environment in which we share and advance together.
No, on the contrary. Open source drives the value of AI & AI engineers down. Companies are only doing it for predatory pricing tactics: drive the value down so no smaller competitors can survive, so you can monopolize the market later.
Looks like on their page they have issues marked as #good-first-issue if you wanted to submit pull requests and help with the code. Or to submit bug reports to help create issues to be resolved.
It comes down to awareness and development. On the development side there's quite a few things that can be done. Open source AI needs more datasets, better labelled datasets, data generation pipelines, and we also need code, implementations, and tools. Can you do Blender well enough to generate trees and grass procedurally? If you look up a RU-vid video, could you learn to? Even if it's not high quality, having automated image generation pipelines could be super valuable going forward. Can you do data labelling? Can you code? Can you develop tools, and extensions? But that's the development side. On the awareness side there's things like Getting the word out there. Creating tutorials and workflows, and sharing them. Gathering lists of assets and keeping tally of performant fine tunes, LoRAs, conditioning (things like libraries of blender images for hand poses), and other useful resources that have a tendency to get buried the week after people first find out about them.
While the scanner is nice, it'd be better if we had some kind of trusted developers list or some kind of verified nodes checkmark. It's way too easy for someone malicious to get their stuff in right now.
This. It's great if you wanna force out your software engineer side, but it's too much. Artists, and folk wanting to play around, arent interested in literally having to learn how this stuff ticks just to even apply, and that's coming from someone who's all about digging in, at least to a degree.
It can be but it also gives you unparalleled granular control and once set up a good workflow can do at once what takes many steps in 1111. It is more complicated but more powerful and efficient in my opinion.
It's just a paradigm that takes a day tops to shift to. All of those inputs/outputs happen under the hood anyway, this way you understand way better what's happening, and everything is more modular.
@@tc8557 You over estimate people's analysis of how much work is involved vs the reward from doing it. The reward really isn't that great compared to others.
Sure but without money these projects do not progress and most users do not financially contribute. These projects cost money and time to produce. People keep making these corporate investors out to be bad guys when they are the reason these project get funded and we end up benefiting.
Nice. Corporations are trying to take and control the coolest thing humanity created. I say fk em. Ai must not be controlled by corporations. Language or generative doesnt matter
I hope it will get as successful as blender org. yo! you said it as I was typing it. haha. also I hope the nodes will look as good as those in blender.
Hm you complain that you can't copyright workflows. Imagine you could. And then think about of ALL the things you wouldn't be allowed to do anymore, because of some greedy businnes copyrighted it.... It would be curse more than a blessing
@@HolidayAtHome Yep it would be like having a box of crayons and someone creates a drawing using a red crayon first and then a blue crayon and then copyrighting that. Nobody else would then be allowed to create a drawing using crayons in that order again. Would never work.
While Open Source is important, the narrative that corporations are the bad guys is flawed. These projects need funding in order to exist and average people don’t really contribute financially but absolutely benefit from these projects. I have zero issue with these investors wanting to make a profit from their investment. People got to eat
Please don't get the ideas confused. Developers are the ones actually making the open source tools. What he's talking about are companies that are taking away the tools from developers and the open source community.
@@AB-wf8ek One is an extension that allows enabling, disabling, and rearranging keywords in the prompts. Beats having to remember what they were and typing them back in. I’ve never heard of that being possible in Comfy
Yeah I'm on it! I developed some simple nodes to help my workflows and they activelly sent a pull request to integrate to the new repository paradigm. That is amazing! I'm going to create even more usefull tools to help comfy go even further l! At least I'll do my best!
I wonder if there could be an initiative to use people’s excess processing power to crunch training data for open source ai. Like SETI at home but for ai.
I hope what you said becomes true. I worked for big corps; can't name because I want to keep myself anon, but it was such bad experience that made me protective about my workflows and methodologies. I will help the community in different ways though.
I would LOVE to see ComfyUI's node based platform extended to LLM's, Agents, Voice, and Music. It needs to move beyond just being platform for Stable Diffusion (which looks to be swirling the drain at the moment).
Free services and products are an opportunity for the company to gather an audience to which it can offer paid additional services. If I just change it to paid, I'll lose the public's trust.
Sounds like a positive move for Comfy. With regards to copyright, I'm sure the artists etc. whose content was taken without their consent for these AI models, would have something to say about that though ...
glad they took the initiative, comfyui is a mess in many aspects. half-backed node engine, half-backed ui, half-backed integrations and nothing really works together.
Thank You for your regular inputs and updates on AI industry. I learned a lot from your videos in last few months. I still learning about AI. I have a one request. Please create a video for Training Lora based on concepts and scenarios. Like a model holding a product say a cup or showing a new book. Also, if you can do a video on merging that LORA in a checkpoint by using super merger it will be a great help. Thanks a lot.
Yea cool, but ummm Many of us dont care to blow lots of extra time trying to do with Comfy what A1111 can do in 10% of the time. However For RU-vidrs who create Comfy How To Videos, this is great..
the problem is that AI relies on data for training, and that data is always "repurposed" from something, so there always is an aspect of stealing, unless the creatives and private persons who produced these data for a specific purpose such as communication or art have declared that they are OK with this data being used for these diverging "AI training" purposes. In AI, the producers of these data are being strong-armed by large corporate "platforms". "Open Source AI" can never be like Linux in this sense - it can never get over the fact that the data used for training is basically - stolen.
Hm not quite the Blender moment I hoped for. Blender Org is completely developing Blender themselfes. Not just interfaces or workflows for 3ds max ... We would need a community funded Base Model with the newest tech present in SD3... and later on the next newest tech. . In the AI there are constantly developments and inotvations. We won't get prompt understanding or text like in SD3 when we forever built on SDXL tech.
We don't compensate the rights holders whose images were used in training our models, so asking for compensation if somebody uses our workflows would be a bit rich. Also very much against the spirit of open source.
Companies can take it, but they can't copyright it either, so if they make something with it that they want to copyright, you can copy it and they can not do anything about it
Even though I've joined `Comfy Org` discord server, I'm not truly sure hwat I've joined to. At a glance, it seems like just any other SD-related discord server. Am I getting right, that it's core feature is that it's an *OFFICIAL* discord server by Comfy devs?
I strongly support this approach. However, as a non-developer user, I think the platform's challenge is to build a compatibility framework that avoids having to manage conflicts between Custom_Nodes, PyTorch versions, etc., which can discourage many users who want a tool "click@work". So, OPEN plateform YES, MESS platform NO. I love ComfyUI
Will it be the same as Stable Difussion, where what was originally open source free eventually becomes paid like SD 3 models. which in the end user is just a testing tool
I totally share your perspective on open source. I think it's important to recognize which projects they are, so we can continue to support them. Thank you!
SideFX Houdini is not open source. It is a proprietary software developed by SideFX, a company known for creating high-end 3D animation and visual effects software. Houdini is widely used in the film, game, and advertising industries for its powerful procedural generation capabilities, robust simulation tools, and flexibility in creating complex visual effects.
Comfyui i think is going to be the next visual programming standard similar to python or js. It may not be the best but if it's what people use it'll be the standard
Open AI not opened. Stability AI not stable..They are failing mission towards of AI to bring human a better life .... I hope comfyUI is really makes us comfy enough.
Commercialization isn't a bad thing. Ideally, something could be contributed back, but at that point it is it's own thing and the open source project can continue to do its own thing. Commercial enterprises can give inspiration to open source projects. The problem would be if a commercial entity interferes with open source via software patent or copyright. So we actually want to embrace pre-empting patents as much as possible. Blocking someone from using open source software as a seed for new enterprises goes against a core open source principle that the user can do whatever they want with the code.
Hyper nerds, FINALLY trying make stuff usable by non-nerds/ nerds who aren't interested in feeling like we're on the clock once we come home!? Finally 😂
Hi Olivio I have a question. I have 2 NVIDIA PC in same home both with low vram (6)(8). Do you know a methode to share the Vram between each other to let them render together?
Every Plattform can be used by people with no coding skills. From the App stores to the web stores. But if something isn't good, people are not going to use it
@@OlivioSarikas Yes, and if your store is full of stuff that doesn't work reliably, because it was made by people who don't know what they are doing, users will stop using your store. As a software developer who tried to use ChatGPT in real projects multiple times, I can tell you that it's simply not there yet. I spent about the same time correcting it's errors than it would take me, to write the code myself.
@@GeorgTomitsch Then don't use it. It works for millions of people. Your concerns are valid, but instead of saying it's doomed, actively work on solutions of how to make it better. Even suggestions help.
You know, that could work. Say you want to set up a web server. All the components, from the networking stack to the front end to the database, all this could be a bunch of nodes you loop together in a UI and hey presto! A web server.
I almost always make them with Midjourney, because it is by far the fastest and easiest for thumbnail generation. i get most of the times what i want on the first roll
Im using comfyui in my new macbook pro m3 pro, can't run any image to image upscaler, 😢 or faceswap! Searching on the internet, can't find the solution yet 😢
Olivio, just regarding: "You cannot copyright methods or ideas." But you can patent it. One of the key innovations Wolfram Research is known for (and patented): "idea of an interactive document that combines code, text, and graphics". Sounds familiar? In saying that, I know nothing about law involving copyright and patents, just thought I'd bring it up. Not sure whether this applies to technologies such as neural networks...
I want a proper NATIVE node that simply lets you switch between inputs. That way you can choose "Img 2 img" or "txt 2 img" or "face restore" or "upscale" or "SDXL/SD3" with "native VAE or Custom VAE" or "LORA" all just rocker toggle switches that look similar to like, a light switch that light up the one you have selected. Also maybe the option to toggle to exclude workflow parts that are disabled.
Efficiency nodes does some of that. You can put the vae, or leave it as baked for the model, add in the prompts, set the clip skip, add loras directly with a drop down, change the way weights are detected, has some hook ups on the back that lets you add other efficiency nodes that let you put as many loras or control nets stacked together as you like, the front end has all the little connections needed for the sample, latent and all that... ONE NODE. Yet, for some reason I never see anyone using it when they make workflows. I'm like... it's a tiny download. Noramlize it, dangit. It cuts out like a dozen other nodes and the advanced efficiency sampler combines several nodes itself. Seriously, three nodes is all you need to render: the loader, the advanced efficiency sampler with or without sdxl tuple, and the save/preview image node which is directly linked to the IMAGE connection on the sampler. ONE CONNECTION no vae decode node, none of that.
ah, the loader also natively lets you choose resolution size and stuff as well. The efficiency nodes have samplers that allow denoise and one that allows you to determine the starting step, so you can refine... in this case, you'd need like five nodes if you want to use two different models. Just 5. Not 5 billion. Yet, no one does work flows with it.
Hello Olivio, thank you so much for your video! I've noticed the same things you mentioned, and I'm thrilled that you're spreading this message, especially as one of the most prominent RU-vidrs in the field of AI generation. Like you, I'm extremely excited about the amazing news surrounding Comfy Org, which is indeed more than just an AI open source project. I'm inspired to continue developing and refining my Comfy automatic workflows, knowing that the platform is in such capable hands.. Thank you for your insightful and engaging content!
Depends on what you mean by "LLM workflow". In theory with ComfyUI you could do about anything that is not interactive and even that is not a strict limitation due to the update-on-change mechanism. In practice there are already nodes to do stuff with LLMs like creating texts or tag images. There is even a node for talking to Ollama which means you can use any model Ollama can run and this locally.
My assumption would be that workflows can be considered as code. It’s a configuration/ code that can be exported and stored into a git. it is a set of instructions that instruct the system to behave in a certain way. Comfy is just the framework which is used in that case. so therefore it should be copyrightable right?
Kind of ironic you would have this concern when the technology we're using is based on data scraped from the internet. At the root of theft is the denial of access. When someone steals your car, the crime is that you no longer have access to your car. If a company makes money off information from the internet, it's not theft, because everyone continues to have access to the same information. If you don't want someone to use your workflow, just don't post it online, but it's not stolen if someone simply makes money from a workflow that was shared online.
@@chrishillery "genuinely curious, I just woke up from a coma, I never heard this term before, I swear! I want you to try to explain a very broad concept to me in very definite outlines"
@eimantasbutkus5324 I've heard the term before, of course. I haven't heard a satisfactory definition though, so I'll often ask someone using it what it means to them.
I enjoyed the conversation. I think it's important to recognize how the open source community is progressing if we want to help sustain it. Belittling people for caring about it seems counterproductive. Just don't watch it if you're not interested.
