People keep stating these look like photoshop filters. This is a very shallow dismissal imo. Why don't you prove it instead by applying some Photoshop filters and achieving a similar effect? Include how long it took as well.
It takes the presenter about seven minutes to go from "image opened in Photoshop" to "comic-style effect applied to image", and that includes a verbal explanation of each step. (The rest of the video talks about things like making sure the image is an adequate size, and adding extra elements like borders and dialogue balloons.)
Thanks for looking. I don't think that effect is similar enough though, but maybe that's just me?
After I wrote my original comment I thought about the filters. If Photoshop had a similar effect available as a filter, I'd assume it was achieved using a similar architecture as the OP presented.
The first animated films to do this were the animated LOTR film and Fire and Ice shortly thereafter, both by the same director.
Logically, very similar during production to how Avatar was made, but doing the rendering "by hand" (and being unable to change the camera after capture).
Pretty impressive results given the budget and time requirements.
Actually while Bakshi was the first to make a movie that went all in on rotoscoping (particularly Fire and Ice where it was 100% rotoscoped while LOTR was a mixed bag) the first animated film to heavily leverage rotoscoping was Snow White.
> The first animated films to do this were the animated LOTR film and Fire and Ice shortly thereafter, both by the same director.
Although rotoscoping is a venerable technique, the results we're looking at with this service don't bear much stylistic resemblance to early works like Fire and Ice or Snow White and the Seven Dwarfs (or more recent rotoscoped works like The Little Mermaid either). If they did, I'd have given them a pass on the 'anime' claim, even though calling the Disney house style — or Bakshi's — 'anime' is still a stretch.
This was built on top of the Pytorch implementation [0] of AnimeGANv2, using gradio [1] and HuggingFace Spaces [2]. See the 100s of examples [3] posted by users.
Nothing is saved by default. Clicking 'flag' saves the input and output (only accessible to the space owner). More info on flagging: https://gradio.app/getting_started/#flagging
Tried a headshot of myself framed like the example pictures and all I got was something that looked like it was bombed by effects out of Paint.NET. Definitely looked nothing like an anime character.
Now that AI figured out how to press the "posterize" button, the singularity is surely just around the corner.
Yeah, glad there were examples, saved me trying it. Results are nowhere close to what I expected. Zero of them looked like anime at all, for one thing. Many truly did look like one or two basic Photoshop filters had been applied, and that was it.
I guess it's impressive you can train a network to do that, but it doesn't look much better than a photoshop filter.
I'm not sure if you're saying it's bad because you're underestimating how powerful some Photoshop filters are, or if you're saying it's good because you aren't.
Nice, but the problem I have with this and similar techniques is that if you zoom out sufficiently far, then you don't see a difference with the original image. This is not how "anime" is supposed to work.
I do second the observation that this isn't nearly "Anime" style by any definition or sample of the style I've ever seen. I would expect to see these results in a GTA 5 loading screen.
Excellent, I was hoping something as accessible as this would appear to play around with this type of model!
Those that dismiss this as nothing more than a few filters are missing the finer details of what it actually does. There is a lot of targeted tweaking, like removing double chins and other old age artefacts. There are small embellishments, like adding a twinkle to the eyes or lip gloss where none was present before. The direction of a gaze is fixed towards the camera. And of course, although it's more cartoon style than anime, the eyes do become bigger.
The quality of your results depends on how appropriate your input is. Try something that has the potential to be Disney and you will get great results.
Sure, you could (with enough practice) get similar results by manually editing a photo. But the beauty of this type of conversion is that it allows for near instant effortless experimentation and (after some other improvements) application to sequenced frames.
The active model does lean towards feminizing and westernizing the subject. Mustaches are removed and an Asian subject will receive wide open blue eyes every now and then.
Viewing these images remind me a lot of caricature portraiture.
As with real artwork, the eye is drawn to embelishments, exagerations, and simplification and directed by composition.
In pieces by a competant artist, these add a new level of interest and dimension to the piece.
With these images, following what makes each image is always a dead end. There is no composition and the embelishments are mechanical, random and uninteresting.
For example, I find myself wondering why the the artist chose to paint the mouth so asymetrical. Is the artist trying to express an emotion? Who is the subject, and why have the been rendered as they have. It's confusing and a challenge to critique the artwork... until I realize its just bit of randomness generated by algorithm incapable of any creativity beyond how the programmer applied generic and randomizing programming methods.
There isn't and cannot be anything unique to any of the images. Once you've studied any one of them, you've really experienced most of what the programmer/artist has contributed.
One big mistake is that this is not anime. This is obviously comic-style, which is more realistic, and, in turn, closer to automated filtering.
Note that Japanese anime normally uses a much simplified style to increase the production speed. This naturally results in a highly distorted(?!) depictions that even humans often find disturbing. This makes anime-style difficult to mimic.
Ugh, it doesn't properly respect EXIF rotation tags so if I try to post from my phone it gets mangled. EXIF rotation has got to be the most absurdly poorly supported feature of a common file type like jpeg.
The example with Bill Gates at the bottom is wild! The "teardrop" under one of his eyes makes it slightly gangster, and the smooth face and awkward smile reminds me of a much younger Bill. Fascinating.
at first the result disappointed me (emotionally, not technically), but after a few minutes I went back to it and was struck by the feel the image brought out. It really reflected some of the wistfulness and sadness I feel with a tiny touch of optimism. I'm not well-versed in anime, but that does seem to be a little bit of the feel that I get from it. bravo.
Not anime _per se_, but the toon-me [0] project gives much more stylistic outputs. I've really enjoyed making some home art with it (and the prior ArtLine).
I found that it did an awesome job on some photos and messed up some. When it messed up, it generally made mouths look way too distorted. I also thought it was funny that it turned obviously green eyes blue. Overall, I think it is a fun implementation, but it would be much better to have a larger output file.
I did my "corporate" picture and it came out looking like I had been badly beaten and with really bad teeth (100 years of coffee and cigarettes). Bit of a blow to the self-esteem.
Who decided this was anime? It doesn't look like any anime I have ever seen, it just looks like a smoothing filter. It looks more like a toonme but not as good and maybe more private? For referece look at this waifu generator, nobody would call these portrait conversions anime. http://thiswaifudoesnotexist.com/
Interesting, you can see how it feminizes faces quite clearly in the examples of Bill Gates and Elon Musk. Gates looks kind of like Ellen DeGeneres to me in the anime version.
This one didn’t do me too bad, but a lot of the “anime filters” I’ve seen will completely feminize my face into an obviously female character despite the facial structure and beard, I assume because I have long hair.
Most of them look like photoshop filter as per other comments, but IU and Billie Eilish definitely looks good.