Crazy AI Animation Workflow - Opus 5.5
Can It Code?
104,510 views
Episode three of building our survival game with AI, and the one I'm most excited about: the animals in the game were never animated by hand. Every hop, every step and every bite comes from an AI video.
We got there in three tries on the same roe deer. First, Opus 5.5 built the deer in Blender from code, rigged it and keyframed a walk. Close, but blocky, and the mesh deforms. Then a much better model from Tripo, with a skeleton the AI built itself and the same keyframed walk: the model was fixed, the motion wasn't. Then the idea that worked: render our own Tripo deer exactly from the side, give that picture to Seedance 2.5 with a strict prompt, and copy the video onto the skeleton frame by frame. A script the AI wrote poses the model, renders it, compares the outline with the video and nudges the bones until they match. 97 frames per clip, about half an hour each, and every video came out right on the first try.
Every animal still came with its own problem: a hare video pretending to be 60 frames a second, deer legs crossing in the gallop, pheasant legs swapping sides and wings Tripo only painted on, and a bear that took GPT-6 Astra fourteen rounds. For each one the AI built a small lab where I go frame by frame and just tell it what looks wrong. Then we gave the animals a tiny brain: two circles, one to stop and watch you, one to run for the trees. The bear doesn't run.
It doesn't work for people yet, and the video shows why: a video is flat, so a single camera can't see depth. We tried a person lifting a stone and walking with a log, and both times the outline matches the video while the 3D is wrong. The fix is more cameras, and that's what we used on our first boss, a troll: Seedance filmed every move from two sides at once. That's the next episode.
Five animals, 21 AI videos and a few days of work. Everything is Godot 4 and Blender, with AI agents working directly on the project files (Claude Opus 5.5, and GPT-6 Astra in Codex). Tripo and Seedance 2.5 (in Dreamina) are tools we used; this video is not sponsored.
CHAPTERS
00:00 Welcome back
00:10 Every move comes from an AI video
00:40 Try 1: Blender, all from code
01:47 Try 2: Tripo + a keyframed walk
02:31 Try 3: an AI video of our own deer
03:15 From video to bones
03:46 All three, side by side
04:00 Every animal, its own problem
04:52 The animal lab
05:07 Any move, with sound
05:29 Animals that run from you
05:51 How their brain works
06:07 Why it fails on people
06:30 A video is flat
06:50 Why the deer works
07:22 Filming from the side doesn't save it
07:42 More cameras, and our first boss
08:04 What's next
CREDITS
Terrain elevation data: GUGiK / Geoportal.gov.pl — Numeryczny Model Terenu. Mapped routes and land areas: © OpenStreetMap contributors — https://www.openstreetmap.org/copyright
#Godot #GameDev #AI #IndieGame #Seedance
20 Comments
Crazy AI Animation Workflow - Opus 5.5
Can It Code?
104,510 views
Episode three of building our survival game with AI, and the one I'm most excited about: the animals in the game were never animated by hand. Every hop, every step and every bite comes from an AI video.
We got there in three tries on the same roe deer. First, Opus 5.5 built the deer in Blender from code, rigged it and keyframed a walk. Close, but blocky, and the mesh deforms. Then a much better model from Tripo, with a skeleton the AI built itself and the same keyframed walk: the model was fixed, the motion wasn't. Then the idea that worked: render our own Tripo deer exactly from the side, give that picture to Seedance 2.5 with a strict prompt, and copy the video onto the skeleton frame by frame. A script the AI wrote poses the model, renders it, compares the outline with the video and nudges the bones until they match. 97 frames per clip, about half an hour each, and every video came out right on the first try.
Every animal still came with its own problem: a hare video pretending to be 60 frames a second, deer legs crossing in the gallop, pheasant legs swapping sides and wings Tripo only painted on, and a bear that took GPT-6 Astra fourteen rounds. For each one the AI built a small lab where I go frame by frame and just tell it what looks wrong. Then we gave the animals a tiny brain: two circles, one to stop and watch you, one to run for the trees. The bear doesn't run.
It doesn't work for people yet, and the video shows why: a video is flat, so a single camera can't see depth. We tried a person lifting a stone and walking with a log, and both times the outline matches the video while the 3D is wrong. The fix is more cameras, and that's what we used on our first boss, a troll: Seedance filmed every move from two sides at once. That's the next episode.
Five animals, 21 AI videos and a few days of work. Everything is Godot 4 and Blender, with AI agents working directly on the project files (Claude Opus 5.5, and GPT-6 Astra in Codex). Tripo and Seedance 2.5 (in Dreamina) are tools we used; this video is not sponsored.
CHAPTERS
00:00 Welcome back
00:10 Every move comes from an AI video
00:40 Try 1: Blender, all from code
01:47 Try 2: Tripo + a keyframed walk
02:31 Try 3: an AI video of our own deer
03:15 From video to bones
03:46 All three, side by side
04:00 Every animal, its own problem
04:52 The animal lab
05:07 Any move, with sound
05:29 Animals that run from you
05:51 How their brain works
06:07 Why it fails on people
06:30 A video is flat
06:50 Why the deer works
07:22 Filming from the side doesn't save it
07:42 More cameras, and our first boss
08:04 What's next
CREDITS
Terrain elevation data: GUGiK / Geoportal.gov.pl — Numeryczny Model Terenu. Mapped routes and land areas: © OpenStreetMap contributors — https://www.openstreetmap.org/copyright
#Godot #GameDev #AI #IndieGame #Seedance
This is far better than those youtubers who just one shot a game from AI and make a slop game.
This game looks like something anyone would genuinely want to play for some time.