VoxelPlot @voxelplot

“Klaus” (Academy Award-nominated) used AI to infer consistent 3D surface information from hand-drawn 2D animation. As Sergio Pablos explains, the key problem is surface correspondence. The system must track the same point on the character across different drawings. This creates something similar to a persistent UV map over 2D animation. Once that exists, textures, lighting and other surface information can follow the character’s volume without using a fully animated 3D model. AI could reconstruct fake 3D passes directly from 2D drawings: • normals • depth • position • material masks • UV / surface correspondence Those passes could drive lighting, shadows, materials and FX while preserving the original 2D animation. Pipeline: ✦ Original 2D animation ⬇ AI-generated textures ⬇ Reconstructed 3D passes ⬇ Lighting / shading / FX It is almost the inverse of Arcane: 3D → made to feel 2D Instead: 2D → given the depth and material richness of 3D What concerns me most is that artists who are open to AI-assisted workflows can face harassment or backlash, making some afraid to collaborate even when they are interested. I really hope this changes over time. I'm particularly interested in workflows involving: • color transfer • rendering and relighting • illustrated texture transfer • AI-assisted compositing • hybrid shots combining hand-drawn animation and generated elements • sequences mixing traditional / handcrafted shots, • hybrid shots mixed fully generative shots in final animation This is the link for the full interview with Sergio Pablos: dailymotion.com/video/x7rcg1l

solar🌙 @solaramv

4 days ago

It probably uses generative AI to map out the topology but man thats just insane

6 2 316 20K 6

@Gabriel_Olson Looks great Gabriel!!

@MetacoGen WOW! This looks awesome. Great work!!

@Mangaka07 You can feed it with naruto scenes!! Just pay attention to camera angles and action

With Wan 3.0 AI anime is getting seriously close to TV quality. You can create sequences like this one that look like an actual seasonal anime. I made a full fantasy fight between a Mage and a Swordswoman — and it genuinely feels anime-ready. Seedance 2.5 keeps characters more on-model. Wan is better at timing, poses and complex layouts. Hailuo H3 sits in between. All three support Vid2Vid and Omni Reference, so you can pass shots and references between them. The possibilities are endless. Interesting times.

@Alibaba_Wan Hi! You liked it!! Do you have CPP program so I can keep making these amazing sequences? 😀

@stellarcbz I don't know exactly, but sround 50$

@rayLuxembourg Mm don't know exactly. Most videos are 2s at 480p

@mooniepush Official webs of Wan, Hailio and Seedance (Dreamlna)

It does not relate what's studios vision about AI and the fact that an individual can create AI Anime alone with a quality enough for people not be thinking only "this looks like AI"; something thay is actually happening. Even on youtube and social media people is starting to get interested in the Universes, stories and characters of solo creators assiested by AI. We are not competing with major studios, we are creating just an alternative to Executive drived enterntainment, where the focus on the MESSAGE is set by the creator itself

@mooniepush I don't use to do it, but I had for specific projects.

@deviljr4 Yes! I am trying ways to control timing better, but the models can't really time frames and poses the same time!

@mooniepush Yes it is!

@macpheeeee I wish only they focused more on characters and story. I am tired of the same "i want to be a hero" highschool boy and goofy artstyle.

@SuperJonBear Just watch a lot of fight anime and take references on how professionals do them, eventually you get the pace, rythm and camera angles.

@PocketScreenAI Thanks John! Appreciate it!

@SilverYogensha Hahaha!! Special touch 😀😀

@mayalinscenes Yeah! I tend to do things that I find interesting for me also and would watch myself too.

In Japanese animation, lipsync does not follow specifc phonems, so you have to fit sentences instead of syllabes. That makes easier to fit good voices, as you only have to stretch frases and, if you have already a gen, you can tag exactly how long it will last: Example: "I think..." (00:00:00-00:00:10); pause 1s, (00:00:10-00:01:10) "..its ok" - (00:01:10-00:01:37) I am using Seedance to generate voices as It has better performance than any voice model tested, by far. To avoid SFX, NO MUSIC, NO SFX and I use a setting in the scene that has no sound. like atmospheric reflexion, off voice, character thinking, quiet empty room, character sitting, etc That way you get a clean voice.

I would suggest once you are building a specifc shot or sequence try every prompt in every model and see how it handles it. Certain things can look better depending on the model. Hailuo is good at replacing a character from a sequence, so if you have a good shot but only the main character has artifacts, I would use hiluo for that. Also, despite bad gens in therms of animation, you can get good layouts, poses and keyframes that you can use then as an input in the model that you find handles best your scene.