AI Experiment — Emotion & Consistency Challenge
I’ve been seeing a lot of AI shorts with fast cuts, big action, dramatic camera moves, and characters flying all over the place. They look cool. But I started wondering: what happens when you ask AI to slow down and actually sit with an emotion? For this test, I wanted to tell a very simple story with a vending machine, a wet little orange cat, some rain, and one warm milk tea in a lonely alley. Sounds simple enough. It was not. 😂 The quiet details became the hardest part. Raindrops changed. Puddles changed. Goggles changed. Bottle changed. Colors wandered off. The camera occasionally developed its own creative ambitions. And timing was a whole other challenge. When a shot is fast, a weird moment can disappear quickly. In a slow emotional scene, every tiny movement suddenly matters. A head turn that is too fast, a light that turns on too early, or a camera move that should not be there can completely change the feeling. AI is very good at making things move. Getting the movement to feel intentional, restrained, and emotional is much harder. There is still plenty I would improve, but I’m happy I kept the slower pace and finished the short. Coming from a traditional CG animation background, I can see the most exciting future in a hybrid pipeline—AI for speed and exploration, with artists keeping control over performance, timing, camera, post, and continuity. Experimented with ComfyUI, Seedance, LTX2, and a healthy amount of editing to convince everything to stay in the same movie.
AI Experiment — Creative Reusable Skill
This one started with just 4 still images and a reusable skill to “what if I loop these together?” idea. No additional software. Complete every step inside MiniMax Design. (The images were loosely inspired by the bright, slightly absurd vacation-photo energy of Martin Parr.) I brought 4 still images into MiniMax Design, used H3 to create a reusable skill for first-frame / last-frame transformation, then generated a 4-second clip, grabbed the last frame, and used that as the starting point for the next one. Then… same skill, next transformation. Repeat three more times until the final shot somehow found its way all the way back to the egg. 🥚 What I really enjoyed was having everything on the Design canvas: reference images, generated images/videos, notes, Agent chat, the reusable skill, and quick timeline edits. No bouncing between tabs, no external links, and much less “wait… which image was I using again? in which folder??” A pretty simple experiment, but surprisingly fun to build.
AI Experiment — Previs Storyboard workflow
Made an animation test of a cat doing very important cat work: knocking things over 😼 Built this using a ComfyUI Previs Storyboard workflow. I still had to partially re-draw 3 storyboard frames for composition, plus some retiming/editing in the final pass — but wow, the speed of getting from idea → almost-final is kind of amazing. Still seeing things to improve, but I’m impressed by how close it got to my scene prompts and reference images. Also really loving the ability to connect multiple reference images for characters + environments. Pretty exciting to see how quickly this workflow can move from a rough idea → previs → finished animation ✨
AI Experiment — MiniMax H3 Handdrawn Live-Action Fusion skill
I think every old bookstore should come with at least one tiny dragon. 📖🐉 Playing with the MiniMax H3 Handdrawn Live-Action Fusion skill and bringing a little magic into the day. ✨
AI Experiment — Video to Video workflow
This time I experimented with a ComfyUI video-to-video workflow by transforming a few simple home videos into a fantasy mushroom village. My biggest takeaway: Getting the right matching frame is the most important part of the process. It isn't just about matching the source video—it needs to communicate the right story and creative intent. I probably spent more time redesigning that frame than anything else. One happy surprise: The guide frame doesn't need to match the original shot exactly. Similar composition and performance are often enough, giving much more freedom when designing the final look. For the ending, I broke my own "matching frame" rule. The giant cat wasn't based on any source video—I added it simply because the story needed one last beat. Sometimes storytelling wins over workflow. Hope those yellow eyes stay with you after the video ends. 🐈✨
AI Experiment — Acting Test
I’ve been seeing so many great Seedance 2.5 acting clips lately, so of course I wanted to try it myself. And then I thought… what if I compare it with another super popular model, MiniMax H3? I kept the setup simple and focused only on performance. The line is always the same: “I’m so happy for you!” I picked 3 acting challenges: - secretive - exaggerated happy - quietly hurt For each emotion, I gave Seedance 2.5 and MiniMax H3 the exact same text prompt and compared how they interpreted the acting. A few things I noticed: Seedance 2.5 gave me some really warm, expressive moments. But the starting frame was less consistent, and sometimes it randomly added extra cups to the table 😂 MiniMax H3 felt more consistent overall, and I liked its follow my text prompt better. The fun part is, now I can keep thinking of more and more ways to say the exact same line with completely different emotions. And unfortunately… “I’m so happy for you!” is now completely stuck in my head 😆 After hearing it this many times, I hope it makes you a little happier too.
AI Experiment — Draw-to-Direct workflow
Been playing with ComfyUI's Draw-to-Direct workflow and I'm pretty impressed by how intuitive it feels. Instead of describing motion with prompts, you can literally draw directions directly onto an image. A few tests I tried: 🐈 Test 1: Draw a path for Object A, and let Object B react to it. A red laser dot becomes the target, and the cat follows it naturally. Not exactly a scientific experiment, but probably the first thing any cat would do. 📦 Test 2: Draw a path that moves behind props while adding camera motion. The cat walks behind a pushcart loaded with storage boxes while the camera tracks the action. 🔴🟢 Test 3: Draw two independent motion paths. The red path starts first. When it reaches the green path, both begin moving together. I like it feels less prompting and more like sketching motion ideas. For decades we've communicated movement through storyboards, animatics, grease-pencil notes, arrows, and timing charts. Draw-to-Direct feels surprisingly close to that language. Not perfect yet, but it's exciting to see AI tools becoming more visual and less dependent on prompt engineering. My biggest takeaway from these tests: sometimes a few arrows communicate an idea more clearly than a paragraph of prompts. Fun times for previs experiments — from idea, to storyboard, to motion. Also, the cats successfully completed every test. woohoo~~