Experimenting with Tech even though I have no clue what I’m doing
Phoenix, AZ
Joined October 2018
- Tweets102
- Following24
- Followers32
- Likes38
Pinned Tweet
Tried making a music video entirely with AI tools. Way harder than I expected, and way better than I expected. More info on how 👇
Pretty random, but Minimax H3 can do stop motion paper animation. Wild what a local model can do. Always finding new things to test
@MiniMax_AI
5090 out of your budget? I made this 5-second clip, sound included, from a text prompt in about 80 seconds on an AMD AI PRO R9700 (~$1,800 vs ~$7,500 for a 5090 right now), running MiniMax H3 in ComfyUI.
The 5090 is about twice as fast, at roughly 4× the price. People say you have to go NVIDIA for AI video. Do you?
Is it as easy as NVIDIA? No. You’ll want Linux with AMD’s ROCm, a ROCm build of PyTorch, and some patience: a lot of speed-ups and custom nodes are written for NVIDIA first.
You’ll hit things that need workarounds or extra memory tuning, and there are fewer guides. But it works, and at a quarter of the price, that’s what counts.
Worked on using a song + extending clips on Minimax h3. I think it came out really good. More info on how 👇
This is 4 clips joined together, the seams are quite difficult to notice and the lip sync stays almost perfect through the whole things. Thanks to Taylor swift for the lyrics haha.
Jargon explanation: the previous clip's last 39 frames are encoded into the latent and frozen with a noise mask, so only new frames get denoised. The song is locked into the audio latent across the overlap and the new part, so the video is generated against fixed audio. Tip: put the seam in a pause. It takes a few frames to re-lock lip sync.
Happy to answer any questions
Genuine question: why is so much AI content surreal? Dreamworlds, robots, dragons.
I've been going the other way, trying to see if AI can capture real, human moments.
Which side pulls you in? Why? Responding to everyone because I want to know.
Here’s my first piece I genuinely worked hard to convey something that feels human. Please share yours.
teapawt.net/jefftriesthings/status…
AI video kept making the small background face twitch during a slow zoom out, so I built a tool to fix it.
It tracks her face through the camera move and locks it in place, frame by frame.
Using my same character for a reimagined taylor swift video. Before vs After 👇
How it works, if you want to try it (Python + OpenCV + ffmpeg):
1. Pick one clean frame where her face looks right, and draw a box around the face.
2. Track the area around that box through the clip (feature tracking) to get how the camera moves, scales and turns each frame.
3. Smooth that motion over time so the tracking itself doesn’t jitter.
4. Paste the face from your clean frame into every frame, warped to follow the camera, with soft feathered edges.
5. Snap it onto the real face each frame (OpenCV ECC alignment) so it sits exactly in place.
The result: the camera still moves, but her face stops twitching.
A lot of tools had to come together. You need a character, then good starting images, a video generator that does lip sync, and of course the actual song too. Consistency with content really is difficult to tackle.
Happy to answer any questions
okay this is actually cool. images, then video, now worlds you can walk around in. i need to try this. now make it run locally lol
Has anyone gotten deep into the agents world? I’ve tested open claw, dots, grok bot. Just not super sold yet. I’d love to be convinced to use them more.
