Spaces:
Running
To my new Model-Mixing Magician!
Hey YnTec,
I've started following you on HuggingFace and I'm totally into your discoveries, your models mixes, and especially... the list of models in your app.py from your "Toy World" Space.
So much so that, since I love programming in my free time and my favorite language is C#, I whipped up a little WinForms project to make it easier for me when I want to generate several images at once. I'm using your models list and the free Inference API service of HuggingFace.
I even found out that during an API call, we can not only provide the "inputs" but also pass an "options" object with 2 boolean properties: use_cache and wait_for_model. The use_cache defaults to true but if set to false, it generates a new image with each call, and wait_for_model is false by default, but setting it to true avoids the 504 error codes (model is loading...) but we have to wait, like... 2 more minutes, that's all!
Anyway, I was wondering, which Spaces should I check out to get your most up-to-date list for the best models :)
And if possible, I'd love to chat (discord or something) with you about how you mix models and other magic tricks you are doing!
Looking forward to it!
Hey BlinkSun! Welcome aboard! It's great to have your feedback, as sometimes it feels kind of lonely here, and I don't even know if I'm going to the right direction, and what I get is confusing, like, getting lots of likes in a model I made... but... it rarely gets used, and then no hearts at all, but 20000 downloads in a week?! o_O Things like that.
Let me get something out of my chest, I think I'm reaching some kind of model saturation, when I started this, or rather, continued from what the giants of hugging spaces had created, I dreamed to be able to provide models that could do everything you could think of! If a single model model can't do it, I could have 5 that put together could get close, and then Dalle 3 happened and it was a shock, I could never get close, no matter what I did, no matter what models I mixed, the outputs just weren't there, and so, I gave up, and compromised, and started delivering the best I could, even if it wasn't what I wanted.
My dream was to get to 1000 models and make a party but what I noticed was that most of the models I was planning to release actually were worse versions of what i already had, or they didn't bring anything new to the table. Civitai previews are tricky as most depend on negative prompts, and those just take a big chunk of the space of possible outputs, so I optimize my models to not require them, but then many, many models become mediocre without them. You may have seen big gaps between one upload and the next, but I'm not on vacation, the other day I tested 11... Yeah, eleven models, and none of them were up to my standards for release, so I spent more days searching until I found one that did it...
Hey, let me show you something from today...
Those are outputs from 4 models I didn't release... so the REASON I merged the Memento model was to have a model to use as a base for Analog Diffusion (https://huggingface.co/wavymulder/Analog-Diffusion) because what bothered me about it is that if you don't put "Analog Style" on it it just has bad outputs... what if we had a model that produced great things no matter what AND those outputs if you added "Analog Style" to it? But no matter what I tried, nothing worked because we already have this:
At the left, we have the output for Analog Diffusion, and at the right, we have Memento's. As you can see, none of my merges really improved over these outputs, she's basically the same as Memento alone, and none of it has the style of Analog Diffusion, so I parked this idea along with many that haven't worked yet because Memento is clearly not suitable for the job...
Anyway, just mentioning it because I really have no one to talk about this, I guess I could have a blog or something, but on the theme of "my failed attempts at merging models" I guess it's better if people don't know about all this, ha!
About the best models: I always try to keep the best one at the top, whenever a model does something that is not up to the standards of Toy World, I sink it deep on the list, sometimes near the bottom. For most recent models you can check this space: https://huggingface.co/spaces/Yntec/PrintingPress - here I try to upload models daily, if I find them, as I become more and more critical about what I upload, and I also alternate with digiplay's models. digiplay is my hero, I learned everything about uploading models from him and the greatest inspiration to merge models was to match his models merges.
About cached models: What kills me is generating images and then losing them before I can save them, that's what clicking Generate does without a cache, and I keep doing it accidentally. With a cache, there's the image again, nothing is lost. Plus, sometimes the errors attack after an image has been generated, but not shown, so clicking generate after such an error restores it from the cache instantly, I can't sacrifice all that for different images from the same prompt.
About wait_for_model: Yeah, I'm using that on my space that allows you to generate with up to 6 models at a time: https://huggingface.co/spaces/Yntec/Diffusion60XX though it's always behind Toy World and the Printing Press because selecting and unselecting models makes the whole list jump around, but one could just select 1 and use it like this, with the cool feature of seeing what actually happens on the errors.
About chatting and Discord: I'm definitively anti-Discord, and in recent times, anti-Private messages and anything that remains concealed from people. I'm glad Huggingface has nothing of that so we have to talk publicly and whatever we talk about can benefit people, but feel free to use this comment section like that, when you post a yellow circle goes around my avatar and I can check what you said quickly.
About secrets: Sure, I'll reveal all my secrets, just beware that they'll be here in the open so everyone will know. Huh, I guess a "secret" is that I could create spaces that allows 6 images at a time with any model, all different, like this one: https://huggingface.co/spaces/Yntec/DreamAnything - or that I have code that would allow you to use negative prompts and adjust the size of your generations so you can have landscape or portraits in your outputs, and 1024x1024 pictures instead of the default 768x768 ones, something I may plan to do for selected models.
And any question you have about merging models, I can answer. I don't have the hardware, so I do all my merging in huggingface spaces, and it takes a while because each image I test takes 15 minutes to generate on there, but I'm not in a hurry and do other things on the meantime. If you want some kind of tutorial about how to merge models and upload them to huggingface I can do that too.
I'm like a music box that you turn once and then I play music for a long time, known for my big walls of text like this one! Hope I didn't forget anything, see you!