@mirthtime

are you having fun in the singularity?

Joined October 2021
  • Tweets2.5K
  • Following710
  • Followers2.8K
356Media
Made a 35s short entirely local on my 5090 with MiniMax H3. Then I told Claude to make it look like an anime animator worked over real footage. It tracked every hit, then added neon ink impact frames, colored speed lines and half-anime cel flashes. The quiet shots stayed clean. The anime FX start hitting at 0:16.
2
182
Apple said we'd see the M5 Studio ultras with 512GB memory in "late october" Any day now?
1
190
I'm obviously going to hell for this, but I thought her video needed more slop.
Udi Wertheimer
@udiWertheimerOct 25
she’s so tired of generative ai
1
1
526
more replies
4/ all the type, captions, chat bubbles and the "Made with AI" watermark that keeps growing are a python + ffmpeg compositor. whisper and demucs handled the timing. the funniest bug: the model kept giving her five fingers, so the six finger hand got its own shot.
1
41
5/ the whole thing was a few hours of GPU time, one clip at a time, and cost basically electricity. yes I see the irony of a song complaining about AI made with AI. that's kind of the joke.
36
I saw @kentdhani's FastH3 Omni piece and wanted to see if I could pull off something like it at home on my 5090. Here's how it was made and where it fell short. 🧵
Kent Dhani
@kentdhaniOct 23
It is FAST. MiniMax FastH3 Omni just went open source, and it can now take up to 9 reference images, video clips, and voice references. Ran it on my RTX5090 in ComfyUI.
2
227
more replies
I never opened an editor. The whole thing came together by talking it through with Grok Bot, about 13 rounds back and forth. But here's where it falls apart. Fast motion turns faces to mush, and can lead to weird blurry artifacts (like you see on the can)
1
44
I tried a face refine pass and it didn't really fix it. So I'd love your take: how are you all handling face breakup on fast motion when using H3? Anyone got a solve for this?
54
Holy moly
Ragavan Thurairatnam
@RagavanThurairOct 23
We made the first AI model you can have a realistic conversation with that was trained entirely on speech! Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs. This is a GPT3 moment for speech.
2
251
Okay, this probably takes the cake for human realism with AI voice. Kudos to @Jack and the team.
jack
@jackOct 23
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. block.xyz/dessa

Dessa - the Conversation Model trained only on speech

block.xyz
103
I spent the day testing every version of MiniMax's H3 video model on my 5090, so you don't have to. Same character, same prompt, same seed. The big difference: some can take a reference photo of your character, and some can only take a starting frame. Here's what that means.
1
1
247
more replies
H3 Fast (officially FastH3 8-Step V2) is a distilled model, like a student trained to copy the full model in fewer steps. About 23 seconds. But it can't take a reference photo. It only takes a starting frame, the first image of the shot, and makes it up from there. That's why his face drifts.
2
91
Links to everything I tested: H3: huggingface.co/MiniMaxAI/Min… Turbo LoRA: huggingface.co/drbaph/MiniMa… FastH3 8-Step V2: huggingface.co/FastVideo/Fas… FastH3 Omni: huggingface.co/FastVideo/Fas… H3 Max: fal.ai/models/fal-ai/minimax…

MiniMaxAI/MiniMax-H3 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co
139
There's also H3 Max. It only runs in the cloud on fal, so you call it through an API and don't need a GPU at all. It's $0.05 per second of video at 480p. I didn't test it here because it can't run on my machine.
1
87
So here's where I landed. Need the same character across shots? Use a model that takes references: Omni for speed, base for max quality, Turbo for quick drafts. Only have a starting image? Fast works. No GPU? H3 Max on fal.
77
H3 Omni (officially FastH3 Omni, the one everyone's posting about this week) is also distilled, but it does take reference images. About 36 seconds, less than half the time of base, and his face held almost as well. That's the sweet spot.
1
69
Load more