Official Hello Wumpus sticker pack from Discord
#stickers #discord #animated #conversion
Add sticker pack to Telegram
Not too long ago Discord added support for stickers. Initially you had to pay for most packs, if I understood correctly. Then they just restricted usage to those with Discord Nitro (like the bs that Durov tries with Telegram emojis). Servers can have stickers, but the number of slots you get depends on server boosts, which requires money, again, if I understood correctly. The good thing about all these moneygrabbing schemes is that Discord created a bunch of official sticker packs of relatively high quality. "Hello Wumpus" is one of these official packs. Wumpus is Discord's mascot or something.
Discord stickers are similar to Telegram stickers in function, but use slightly different formats and have different restrictions. Vector animated stickers use the same Lottie technology that Telegram uses, but instead of gzipped 512x512 animations with vague complexity restrictions, Discord's stickers use 320x320 uncompressed animations without complexity restrictions. So in order to convert the files, I wrote an app which performs the transformation directly on Lottie JSON. I had to fall back to the long conversion process with After Effects in a few cases though because Telegram complained about the complexity of 10 masks in a simple sticker.
Please open Telegram to view this post
VIEW IN TELEGRAM
âĪ6âĄ4ð2
#dev
This sticker may look simple, but it isn't. Notice how the shape of every limb changes with rotation, space suit wrinkles are perfectly aligned with limbs, the air tub is attached to the body and tail changes its shape with the strength of wrinkles depending on the angle of bending.
I found a way to do time remapping/stretching in vector stickers which isn't supported in Telegram, and to do wiggle path which isn't supported even by the full Lottie library. It took a lot of expressions to get there, but it's possible!
Every limb consists of base shape, shared contour and extra lines on top of body which visually separate them from the body. In case of tail, it's 4 separate path-animated pieces. Each is animated from 0:00 to 0:01 from one extreme position to the other, then everything in between is fixed to properly align. Then each animated property (position, rotation, scale, path) is mapped from an expression control value to time, with a controller for every animated piece.
This not only allowed time stretching, but made adjusting interpolation of path keyframes as simple as moving a Bezier handle, without dealing with any stupid speed graphs where making the first derivative continuous is close to impossible.
As a bonus, stars are animated with an expression on their paths which wiggles each vertex and handle separately, with controls for frame hold duration, amplitude and frequency.
What's the price of this? Expressions must be converted to keyframes, otherwise they won't work in Telegram. That means that you get a separate keyframe for every single frame. Each animated position transform contains 180 keyframes, each path-animated shape contains 180 separate shapes (3 seconds at 60 fps). This grows fast, it's super easy to reach the limit of 64 KB.
My optimizing app somehow reduced 14 MB of the final JSON into 64 KB, but it was very tight and I had to reduce precision a tiny bit. Using this technology in any actually complex sticker would be impossible. I need to work on optimizing compression of path animations. Something as simple as approximation with linear interpolation across all vertices may probably reduce data size by 80%.
I'm also curious on how this can be used in DUIK's limbs animated with inverse kinematics, as technically it should allow changing the shape of muscles depending on the angle, fixing joints at extreme angles and things like that. I don't need it often as a lot of stickers don't have a lot of extreme movement in them, but it can be very useful in a few cases. The remaining question is how to plug the path remapping math into DUIK's expressions.
Please open Telegram to view this post
VIEW IN TELEGRAM
ðĪŊ5âĪâðĨ3ð2ð2
Please open Telegram to view this post
VIEW IN TELEGRAM
Disney: The Lion King
#emoji #disney #lion #feline #movie #conversion
Add emoji pack to Telegram
The most important pack from Disney Emoji Blitz! As usual, with two extra emotions for the main emojis, compared to the original, with all the props and extras I could find in the resource files.
Bonus: The Lion King as Told by Emoji video from the creators of Disney Emoji Blitz. Better than The Lion King 2019, isn't it?
Please open Telegram to view this post
VIEW IN TELEGRAM
âĪâðĨ5âĪ4ð4ð1ðą1
After taking FOREVER to improve the technology, I'm catching up with the current commissions. Long story short:
1. Contour splitting and scripting: DONE
2. Soft bending of vector images along Bezier curve: DONE
3. Integration of time expressions with DUIK: DONE
4. Keyframe trimming: BUGGY (?)
5. Path keyframe optimization: PARTIAL
6. Spatial bending like AE pins, but for vector images: IN PROGRESS
7. Mask-less masking with rounded edges: PLANNED
8. Learning the potential of diffusion models for stickers: DONE
What this means here and now: I can do tails of any complexity, with any patterns, with any reasonable bending. So the tail I did for the wagging snow leopard sticker won't take me 6 hours to animate AND it'll look much better, without any glitches. This has been bugging me since forever, as I want tails that look like tails and not like sticks.
I can bend limbs of any complexity at any angle. Previously, any sort of multi-colored patterns around joints would require sacrifices. Contours and overlays tech allowed making contours look like contours instead of integral parts of the fixed geometry and allowed putting various patterns on top of joints. Now I can also fix inner lines which were previously split by rotation. This thing is much harder to notice and not every sticker heavily relies on it; it's rather a sort of thing which you notice only when it's missing.
Also, on the whole diffusion kind of machine learning boom of late. While getting a decent looking image is possible with both open and closed models, getting the overmind to understand multiple characters, complex descriptions, specific designs, face expressions, non-trivial poses and a myriad of other concepts is close to impossible with the current publicly available tech. I hoped I'll be able to produce some sort of ML-assisted stickers for meme purposes, but looks like I'll have to wait before the tech gets to the point when it's possible.
1. Contour splitting and scripting: DONE
2. Soft bending of vector images along Bezier curve: DONE
3. Integration of time expressions with DUIK: DONE
4. Keyframe trimming: BUGGY (?)
5. Path keyframe optimization: PARTIAL
6. Spatial bending like AE pins, but for vector images: IN PROGRESS
7. Mask-less masking with rounded edges: PLANNED
8. Learning the potential of diffusion models for stickers: DONE
What this means here and now: I can do tails of any complexity, with any patterns, with any reasonable bending. So the tail I did for the wagging snow leopard sticker won't take me 6 hours to animate AND it'll look much better, without any glitches. This has been bugging me since forever, as I want tails that look like tails and not like sticks.
I can bend limbs of any complexity at any angle. Previously, any sort of multi-colored patterns around joints would require sacrifices. Contours and overlays tech allowed making contours look like contours instead of integral parts of the fixed geometry and allowed putting various patterns on top of joints. Now I can also fix inner lines which were previously split by rotation. This thing is much harder to notice and not every sticker heavily relies on it; it's rather a sort of thing which you notice only when it's missing.
Also, on the whole diffusion kind of machine learning boom of late. While getting a decent looking image is possible with both open and closed models, getting the overmind to understand multiple characters, complex descriptions, specific designs, face expressions, non-trivial poses and a myriad of other concepts is close to impossible with the current publicly available tech. I hoped I'll be able to produce some sort of ML-assisted stickers for meme purposes, but looks like I'll have to wait before the tech gets to the point when it's possible.
ð18âĪ5ðŊ2
Removed all spam followers on my Twitter account https://twitter.com/CrazySnep. Went from 40 subscribers to 4. Feeling successful. ðĪŠ
The hell are all these accounts with 1000 follows, 10 followers and 3 random photos of half-naked ladies?
The hell are all these accounts with 1000 follows, 10 followers and 3 random photos of half-naked ladies?
ð13âĪ7âĪâðĨ2