Uploaded January 2025 | Updated November 2026, 5 days ago
I’m diving into some of the most exciting breakthroughs in AI video generation. First up is a mind-blowing technique called Diffusion as Shader (DAS), which merges 3D tracking with diffusion so you can seamlessly control subjects, locations, and even camera movements. Then, we explore two new game-changers from Adobe—one of them is open-source!—including TransPixar for text-to-video with built-in transparency and FaceLift for single-image 3D head reconstruction using Gaussian splatting.
I’ll also give you a rapid rundown of cool new updates from MiniMax, Runway, and Krea, like subject reference tools, 4K upscaling, and built-in audio generation. You’ll see how these tools push boundaries, from full-body animations to real-time 4D facial capture.
LINKS:
Diffusion As Shader: igl-hkust.github.io/das
TransPixar: wileewang.github.io/TransPixar
Facelift: wlyu.me/FaceLift
Chapters
00:00 – Intro
00:30 - A Breakthrough in AI Video (DAS)
00:51 - A Brief Primer on Diffusion and Shaders
01:59 – How DAS Solves 2D Animation Issues
03:08 – Benefits of 3d for AI Video
03:44 - Using DAS with FLUX Depth
04:22 – Video-to-Video Magic & Mesh Animations
05:26 – Adobe’s Open-Source TransPixar for Transparency
07:22 – FaceLift for 3D Heads & 4D Novel View Synthesis
08:44 –MiniMax Subject Reference Now Available and What's Next
09:15 - Gen-3 4K Upscale
09:46 – Krea's Video and Sound Generation
I’m diving into some of the most exciting breakthroughs in AI video generation. First up is a mind-blowing technique called Diffusion as Shader (DAS), which merges 3D tracking with diffusion so you can seamlessly control subjects, locations, and even camera movements. Then, we explore two new game-changers from Adobe—one of them is open-source!—including TransPixar for text-to-video with built-in transparency and FaceLift for single-image 3D head reconstruction using Gaussian splatting.
I’ll also give you a rapid rundown of cool new updates from MiniMax, Runway, and Krea, like subject reference tools, 4K upscaling, and built-in audio generation. You’ll see how these tools push boundaries, from full-body animations to real-time 4D facial capture.
LINKS:
Diffusion As Shader: igl-hkust.github.io/das
TransPixar: wileewang.github.io/TransPixar
Facelift: wlyu.me/FaceLift
Chapters
00:00 – Intro
00:30 - A Breakthrough in AI Video (DAS)
00:51 - A Brief Primer on Diffusion and Shaders
01:59 – How DAS Solves 2D Animation Issues
03:08 – Benefits of 3d for AI Video
03:44 - Using DAS with FLUX Depth
04:22 – Video-to-Video Magic & Mesh Animations
05:26 – Adobe’s Open-Source TransPixar for Transparency
07:22 – FaceLift for 3D Heads & 4D Novel View Synthesis
08:44 –MiniMax Subject Reference Now Available and What's Next
09:15 - Gen-3 4K Upscale
09:46 – Krea's Video and Sound Generation


![Octo Is Seedance’s Prompt Killer
Dreamina just rolled out Octo, a new canvas/agent workflow that tries to turn Seedance 2.0 into something closer to a scene builder than a prompt box.
So I tested it with character sheets, storyboards, a cyberpunk fight sequence, and a timeline extension to see whether Octo actually makes AI video workflows easier, or just moves the chaos into a new interface.
We’re also looking at NVIDIA Cosmos-3, Black Forest Labs showing Flux to Martin Scorsese, Bernini as an open-source “Google Omni for video” style model, and TripoSplat for turning 2D images into controllable 3D-ish spaces.
Basically: AI video is getting less like “write one prompt, get one clip” and more like planning, editing, world-building, and trying to keep the machine pointed in the right direction.
Big thanks to #genspark for Sponsoring Todays Video. You can get Free Credits at Genspark by using this link: https://www.genspark.ai/?utm_source=yt&utm_campaign=TheoreticallyMedia02
#WorkWithGenspark
Links:
Dreamina / Octo: https://dreamina.capcut.com/ai-tool/home?utm_source=Officiaaccount&utm_campaign=octo&utm_content=6ytb
NVIDIA Cosmos-3: [add link]
Cosmos-3 on FAL:
Black Forest Labs / Scorsese clip: https://x.com/bfl_ai/status/2061800769226252721?s=20
Bernini: https://bernini-ai.github.io/
TripioSplat: https://huggingface.co/spaces/VAST-AI/TripoSplat
CHAPTERS:
00:00 AI video workflows are heating up
00:24 Dreamina launches Octo
01:01 Using Octo as a canvas agent
01:51 Character sheets and storyboards
02:28 Where Octo starts to struggle
03:14 The Seedream / image model problem
04:03 Testing Octo with a cyberpunk character
04:28 The agent brain feels underpowered
05:27 Fixing the character sheet manually
05:56 Building the scene anyway
06:58 Turning the elements into video
07:35 Octo’s timeline extension
08:27 The part Octo actually nails
08:38 Is this the future of AI video workflows?
10:16 NVIDIA Cosmos 3
11:38 NVIDIA wants to be the physics department
12:21 Martin Scorsese meets Flux
13:03 Sponsor - GENSPARK
15:21 Bernini, an open-source Omni-style video model
16:23 How Bernini plans video edits
17:17 The H100 problem
17:49 Triple Splat turns images into 3D spaces
18:32 The cheap World Labs workaround
19:07 Wrap-up
#seedance2, #nvidia, #opensourceai Octo Is Seedance’s Prompt Killer](https://i.ytimg.com/vi/EmzS3MiXLfU/mqdefault.jpg)







