Honestly the more cartoony style actually makes your version better than apple's. Because its simpler there are less errors, the flying pig from apple has a slightly weird look to it while yours looks natural
oh wow, seeing an open-source project created in much less time beat a product which was being worked for years by a company with an almost infinite budget is insane. Awesome work.
its more resource intensive and is using a already available model with a lora, no model was trained or developed. Its still a nice video but a misleading title
What I think is also really stellar about this is how, at least partially, this process is ethical. For Apple, they are training on their emoji’s specifically. For you, you’re training on what is almost (if not literally) Creative Commons images. It’s so cool to see that with the valid concerns surrounding how this fun technology has been used
While the emojis generated by flux aren't quite at the level of apples, this is extremely impressive for something that was trained on one laptop in 10 hours.
1:39 hate to break it to you but by viewing the index page you are making the hundreds/thousands of requests you tried to avoid. After all, the images have to be shown somehow on your device. Web servers are designed to handle this kind of load and unless you’re hammering the server 24/7 or from multiple devices at once, there should be 0 issues. But really cool project!
I still found it cool that he intentionally tried to avoid spamming the web server when scraping, this is something I think not a lot of people seem to be aware of.
Two easy improvements that should be applied: Resolution matters alot when generating images with sd/dit/flux. Sinсe models were trained in way higher resolutions (512, 1024), you will get way better results if you generate your images at that "reference" resolution. Second: use BLACK background, not white. Objects on sides are darker because of light falloff. And most of real object has no white halo around them. So, just use black bg. Also you could use another model for removing bg, like Inspyrenet.
*Background Music in this video* 0:00 - 2:56 - LAKEY INSPIRED - The Process 3:01 - 4:09 LAKEY INSPIRED - Warm Nights 4:13 - 4:51 - Happy High - PRNV 4:59 - 6:21 - Faithful Mission - Artificial.Music 6:24 - 8:43 - Home - We're Finally Landing Even if these don't need credit and are copyright free, some people might want to use them for their own work or just listen to them! I've never done a music list comment like this before but I regonized some songs so why not put them here. If you know of the others then reply EDIT: He added them in the description too! I commented this before he added it but I'll keep this comment here too I guess
It’s funny looking at giant corp haters in the comments not understanding that genmoji runs offline on a mobile device. This project even though it’s impressive, it’s just a lora for flux (which is released by another company) is too resource intensive (on a 3080ti 12gb it would take about a minute to create an image). This project by no means beats what Apple achieved.
Just further proves that Apple Intelligence doesn’t require an iPhone 15 Pro or newer. If anyone can recreate it they are purely saying it’s a RAM issue to sell new iPhones.
Flux is an awesome model to start with. A comparison between your trained model and Flux w/out training would be awesome (using the same LLM reprompting)
8:16 is quite literally Bojack Horseman. Tbh, I like your version of the flying pig better than Apple's. Awesome work, man! I'll tinker around with it for sure.
You should Look at "Transparent Image Layer Diffusion using Latent Transparency" for a Stable Diffusion model that handles transparency out of the Box.
Looks good already. If you add actual Genmojis to your dataset, it would have immediately produced a LoRA that looks more like Apple's. If you have enough Genmojis it may even learn the aspect of merging traits from different emojis into one. I can only assume that Apple did something similar. Also, just the few word emoji description is likely not enough to ensure the LoRA is able to pick up certain character traits to merge them with other emojis.
@@Gorilla_Snifferalso the apple emoji files have a weird license, tldr youre not really allowed to use them if youre not prompting for apple or for the apple platform like an app. Though I never heard that Appl actually goes against it but then using these files as training data, yeah I dont know. Google Noto Emojis or even Microsoft Fluent Emojis are free to use.
Open genmoji looks more apple than apple genmoji itself Maybe because you have a entire prompt while apple just has "Paraphrase this" (at least im pretty sure based off the video)
I highly disagree, the Apple Genmoji is an unique style that reflects the Apple culture very well. The open Genmoji just reflects the existing emoticon style of, mainly, apple devices but mostly the general emoticon style of all devices. There's a big difference and the Apple Genmoji is a million times better than this implementation.
I would try adding a final feathering step to help get rid of the white outline. 2-4 pixels should be plenty. It would also make it softer and help with that more rounded Apple design.
Can you try to run the base (not fine tuned) model with your generated prompts to see if fine tuning meaningfully improved the output? I’d assume flux has seen every emoji there is during its training.
4:00 About the GPU overheating-this is probably due to Apple boosting power performance without upgrading the cooling system. Makes me wonder how these would perform on an M1 Max (I have the M1 max, but zero coding experience. HMU if u want to do a comparison)
Genmoji gives too detailed outputs, which is not fit for the emoji style. Its supposed to be simple and visible in a small scale. Your genmoji nailed it!
Many different CPU types now have NPUs, unless they are EXTREMELY underwhelming, IDK why you would be restricted to a machine with a beef cake GPU if there are options for NPUs available
Unsure why it was seen as such a problem describe the emojis as that's what the website's already doing it already downloads all 5000 or whatever of the emojis just like you would? So surely the server can handle it
Honestly the more cartoony style actually makes your version better than apple's. Because its simpler there are less errors, the flying pig from apple has a slightly weird look to it while yours looks natural
not to mention the robber with a money bag example where apples model drew up some nightmare fuel
The popsicle was my favorite, felt more "Appley"
Weakest open source chad vs biggest company
Long live the world of FOSS!
@@MoolsDogTwoOfficial LONG LIVE FOSS!
Weak chad
Huh
oh wow, seeing an open-source project created in much less time beat a product which was being worked for years by a company with an almost infinite budget is insane. Awesome work.
beat?
Years?
its significantly more resource intensive
its more resource intensive and is using a already available model with a lora, no model was trained or developed. Its still a nice video but a misleading title
beat? did you see how it looks like? and it takes much more performance
What I think is also really stellar about this is how, at least partially, this process is ethical. For Apple, they are training on their emoji’s specifically. For you, you’re training on what is almost (if not literally) Creative Commons images. It’s so cool to see that with the valid concerns surrounding how this fun technology has been used
Honestly I prefer the emojis outputted by flux more than Genmoji in most instances. Nice work!
true that
Interesting how apple’s version of “robber with a money bag” turned out to have a black skin tone.
@@GordonChil if you notice that and think it has actual meaning, you're more of a racist than anything
@@GordonChilthe “robber” is just another bag tho? It’s not a person 😭
I like that your emojis are less detailed because the original ones are simple as well
While the emojis generated by flux aren't quite at the level of apples, this is extremely impressive for something that was trained on one laptop in 10 hours.
1:39 hate to break it to you but by viewing the index page you are making the hundreds/thousands of requests you tried to avoid. After all, the images have to be shown somehow on your device.
Web servers are designed to handle this kind of load and unless you’re hammering the server 24/7 or from multiple devices at once, there should be 0 issues.
But really cool project!
well said.
I still found it cool that he intentionally tried to avoid spamming the web server when scraping, this is something I think not a lot of people seem to be aware of.
definitely got to respect him for caring about the load he puts on that site's server
Yep and it’s guaranteed that it’s all cached with a CDN anyway probably across nodes around the world 😂
@@Conclusions that's a big assumption for a random website about emoji's.
8:26 lol
Apple genmoji: actually racist
What, brown bags always take money
@@SalzmanSoftwareand Apple’s version doesn’t even have a person (robber) lol 😭😭
I like how it turned the robber INTO the money bag in some "The Fly" type stuff
Two easy improvements that should be applied:
Resolution matters alot when generating images with sd/dit/flux. Sinсe models were trained in way higher resolutions (512, 1024), you will get way better results if you generate your images at that "reference" resolution.
Second: use BLACK background, not white. Objects on sides are darker because of light falloff. And most of real object has no white halo around them. So, just use black bg. Also you could use another model for removing bg, like Inspyrenet.
*Background Music in this video*
0:00 - 2:56 - LAKEY INSPIRED - The Process
3:01 - 4:09 LAKEY INSPIRED - Warm Nights
4:13 - 4:51 - Happy High - PRNV
4:59 - 6:21 - Faithful Mission - Artificial.Music
6:24 - 8:43 - Home - We're Finally Landing
Even if these don't need credit and are copyright free, some people might want to use them for their own work or just listen to them! I've never done a music list comment like this before but I regonized some songs so why not put them here. If you know of the others then reply
EDIT: He added them in the description too! I commented this before he added it but I'll keep this comment here too I guess
Thank you for your service.
I appreciate this!
Thank you!
sorry to tell you but he already credited it in the description
@@Bird526 That wasn't there when I commented this, he added that afterwards I think
the robber and the teddy bear space suit examples you showed are done better by Open Genmoji than they are by Apple!!
Bro, your emojis actually look more apple than genmoji, give yourself more credit
Just a thought, maybe you can use the injected prompt from the note app to rephrase all of the emojis names in the training process.
200 iq moment
It’s funny looking at giant corp haters in the comments not understanding that genmoji runs offline on a mobile device. This project even though it’s impressive, it’s just a lora for flux (which is released by another company) is too resource intensive (on a 3080ti 12gb it would take about a minute to create an image). This project by no means beats what Apple achieved.
Just further proves that Apple Intelligence doesn’t require an iPhone 15 Pro or newer. If anyone can recreate it they are purely saying it’s a RAM issue to sell new iPhones.
Flux is an awesome model to start with.
A comparison between your trained model and Flux w/out training would be awesome (using the same LLM reprompting)
Loved this style of video, you got yourself a new sub🔥🔥🔥
Love how Google has something so much better than Genmoji IMO without using any AI.
What do they have? I haven't seen anything from Google?
@@consolegamer1122he's probably talking about Google's Emoji Kitchen
@@consolegamer1122 It's called Emoji Kitchen. They had it for years. You can look it up and try it in the search or just use it on Gboard.
@dogo231fr Exactly. Has been there for years, and it‘s integrated in GBoard.
@@juniorwmg Apple got that too
8:16 is quite literally Bojack Horseman.
Tbh, I like your version of the flying pig better than Apple's. Awesome work, man! I'll tinker around with it for sure.
You should Look at "Transparent Image Layer Diffusion using Latent Transparency" for a Stable Diffusion model that handles transparency out of the Box.
Looks good already. If you add actual Genmojis to your dataset, it would have immediately produced a LoRA that looks more like Apple's. If you have enough Genmojis it may even learn the aspect of merging traits from different emojis into one. I can only assume that Apple did something similar.
Also, just the few word emoji description is likely not enough to ensure the LoRA is able to pick up certain character traits to merge them with other emojis.
I wanted to do this but couldn’t figure out how to approach it. I can’t wait to see how you attack it.
yea getting the emoji files and trying to create an image based on the prompt whilst using the emojis was such a hassle and eventually I just quit.
@@Gorilla_Snifferalso the apple emoji files have a weird license, tldr youre not really allowed to use them if youre not prompting for apple or for the apple platform like an app. Though I never heard that Appl actually goes against it but then using these files as training data, yeah I dont know. Google Noto Emojis or even Microsoft Fluent Emojis are free to use.
1:42 it's not racism, it's optimization!
>I don't want to overwhelm their serv-
Erm xir that's what your browser does and it seems fine
wow this is awesome! you've inspired me to try this out, amazing work man
Nice work, this is amazing!
Open genmoji looks more apple than apple genmoji itself
Maybe because you have a entire prompt while apple just has "Paraphrase this" (at least im pretty sure based off the video)
I highly disagree, the Apple Genmoji is an unique style that reflects the Apple culture very well. The open Genmoji just reflects the existing emoticon style of, mainly, apple devices but mostly the general emoticon style of all devices. There's a big difference and the Apple Genmoji is a million times better than this implementation.
@@Melvin420x12ditto
@@Melvin420x12 ai reply lol
Nah apple looks more apple
@@Melvin420x12wtf is even "apple culture", they are just making expensive devices.
Great work!
I would try adding a final feathering step to help get rid of the white outline. 2-4 pixels should be plenty. It would also make it softer and help with that more rounded Apple design.
You could use layer diffuse for the transparent images
Dude your OpenGenmoji is better than Apple's. If i were you and i had budget and time I would set up a server and make an app for OpenGenmoji.
Can’t wait to test it out
"I painstakingly wrote a Python script"
*proceeds to just tell an LLM to write it*
This is awesome bro!
3:50 it started generating goatse images lmfaoooo
to get the emojis name i would just use shortcuts and use the action "Get Name of Emoji", it will show you official names of the emojis
I think u can also just select the emoji, and instead of copy or paste, keep swiping until you see speak, Siri will say the name of the emoji
8:21 open genmoji looks more apple than apple
Amazing and honestly better actual cartoon style 😎🎉👍
8:23 Open gen > Apple gen here
Open gen: robber with a money bag
Apple gen: bag with face next to smaller money bag
This was worth watching
Can you try to run the base (not fine tuned) model with your generated prompts to see if fine tuning meaningfully improved the output? I’d assume flux has seen every emoji there is during its training.
4:25 I actually thought yours look more emoji-ish than Apple's !
How would i be able to intergrate this with XCode?
You should try and train a model that can fit on an iPhone like genmojj
And he even did it better!
I'm speechless!!!
3:46 just a casual 180° at step 1600
😧
izlediğim keyifli videolardan biriydi...
hey so this is actually really good!
The summoning salt music hahah so fitting tho
did you re-train your model after generating descriptions from the llm?
Please, can you give me the dataset?🥺
Awesome work!
you could try to make a genmoji site so you can make ai emojis there two
Yours looks so much cuter too
Maybe I overlooked it but what LLM did you use to generate the prompt?
Very cool video and project btw!
If you notice it looks like apple exaggerate their shadows on the emojis, maybe you should add that
Your results are mostly better than theirs. Wild.
4:00 About the GPU overheating-this is probably due to Apple boosting power performance without upgrading the cooling system. Makes me wonder how these would perform on an M1 Max (I have the M1 max, but zero coding experience. HMU if u want to do a comparison)
Open genmoji looks.. better.. wow!
8:25 You atuallly made it better then apple... :D
What font in vscode are you using?
This is insanely underrated.
i thought you were gonna remake the emojis genmoji makes without ai to spite ai or something
8:27 Its crazy that apple made the robber black
Crazy work
gg bro !!! such a great work fr
This is so good
"The og skin tone" is fucking wild xD
Bro has a $4000 Laptop and thought we wouldn’t notice 💀
Genmoji gives too detailed outputs, which is not fit for the emoji style. Its supposed to be simple and visible in a small scale. Your genmoji nailed it!
Why white and not some easier to remove colour?
we should use this to make the seahorse emoji
Isn’t there a website that allows you to combine almost any two emojis
Emoji kitchen by google
isnt this just the emoji kitchen
8:27 is that a racism
No
@hanmira thanks for clearing this up
Basketball man takes money
Nice, been meaning to do this with my discord emotes lol
BRO I JUST REALISED THUMBNAIL IS TECHNOBLADE REFERENCE
I can’t open the project
Wow really nice! 🌟
U deserve more subs
Wow, great project!! I'm curious if you tried using the Unicode CDLR to get the emoji names?
"painstakingly" Github Copilot
3:10 Diese Kommentarsektion ist nun Eigentum der Bundesrepublik Deutschland.
Google Emoji kicken!
Many different CPU types now have NPUs, unless they are EXTREMELY underwhelming, IDK why you would be restricted to a machine with a beef cake GPU if there are options for NPUs available
really goes to show that the only thought apple put into it was "Oh hey guys, ai"
Actually, apple should make this official.
8:24 why did they make the robber black xDD
8:25 ok but I hate how Apple sometimes does dumb stuff like this 😭 like what was the ai thinking 😭😭
You did all this on a MacBook? Wh- how? How is its GPU adequate for this purpose?
Your emojis look better/more emoyi like than apples i think
only for apple? okay
8:30 this is pretty comparable
why is there a background
Combine two lora, preferably a Disney 3D character style lora
Of all the “intelligence” features that would be nice this isn’t one of them.
Unsure why it was seen as such a problem describe the emojis as that's what the website's already doing it already downloads all 5000 or whatever of the emojis just like you would? So surely the server can handle it
Awesome.
looks great some are better lookin than apples and some are worse but cool made :)