This episode runs from a $70 million research bet on A24 to a run of open-weight model releases and one big acquisition. We picked apart what that check actually buys, then worked through open weights you can fine-tune at home, trainable video effects, and a familiar upscaler changing hands.
Jump to:
01:11 Fable's one-weekend release
05:40 The $70 million A24 research deal
17:01 Krea 2 goes open weight
21:48 LTX Trainer and trainable video effects
27:03 A veteran upscaler gets acquired
Quick Take
Two threads ran through this one. First, Google DeepMind putting money into A24 while veteran directors line up behind AI tools and newer ones push back. Second, a set of open-weight releases that move model training out of the big labs and onto your own machine. We kept circling the same point. Film is small money for Google, so the real prize sits somewhere past the movie.
What We Debated: What Google's $70M Actually Buys Inside A24
We spent the segment pushing past the press-release version. The deal puts $70 million into A24 through a funding round, with Google as an AI research partner, and the framing we read is that Google is not licensing A24's catalog and not training on its films. The stated aim is building tools for the filmmaking process, and the reported terms (put at $75 million in early coverage) lean on the safe examples.
The comfortable framing. Previs, prep, and storyboarding are the uses everyone leads with because they draw the least heat.
Not Google's first film move. It already worked with Darren Aronofsky and sat as a tech partner on Doug Liman's Satoshi film, and Martin Scorsese has now signed on with Black Forest Labs' FLUX, also pitching storyboards and previs.
The generational split. One of the greatest living directors is embracing the tools while an emerging filmmaker like Kane Parsons says he would never touch AI.
Joey's argument: Google's interest is not the box office. Film is tiny next to gaming, robotics, and self-driving cars. A24 is a way to feed and stress-test world models like Omni, and a trusted studio partner could get a less-restricted build that yields better training data. Google's video ad network and YouTube creation tools also need strong video models.
Addy's counter: the data play is the elephant in the room. A24's highly stylized, intent-driven content is the highest-grade training data you can get, far above stock or open-web material, so the clean "no data" line is hard to fully believe. He read A24 as doubling down to survive against AI-native studios before one grows into its place.
Addy drew a two-year timeline and put Google's motive plainly:
Getting their video models to behave better, more accurately, have better control, it just feeds into their overall AI play of training robots, training cars, delivery vehicles. All of their big industrial stuff depends on having high-quality world model, video model, image model. And so that's their play.
The YouTube tension complicates it. Google builds AI creation tools, then shuts down channels accused of leaning on AI too hard, including creators who say their work was fully animated in Blender. Where the line sits between encouraged and too much is still unclear.
What We Flagged: Fable's One-Weekend Release and Midjourney's Body-Scan Pivot
Anthropic released Fable, a safer cut of its more restricted Mythos model. It was live for a weekend, and it ran agentic coding tasks with almost no supervision. We watched it build a working app on its own for close to an hour before a Trump-administration order restricted it to US citizens only, which Anthropic had no way to enforce, so the model came down.
The image model Midjourney also moved into biomedical imaging. Per the article we discussed, ultrasonic body scans arrive as heavy acoustic noise, close to the Gaussian noise that image diffusion already learns to remove, so the same denoising approach can reconstruct the scan. The pitch is cheaper, more frequent scanning for earlier detection.
What We Explored: Krea 2 Open Weights and Why LoRAs Are Back
After a closed Krea 2 image model, the open-weight version shipped in two pieces: a larger raw model for fine-tuning and building LoRAs, and a turbo model for fast local inference. It runs locally, plays well in ComfyUI, and lands close to real-time.
A few points we kept coming back to:
- Looks like text-to-image only. It appears to have no image-to-image or edit mode, so it fits still imagery, design, and mockups rather than character sheets or first frames.
- Trained on the ugly stuff. Per Krea's technical report, the model learned on a mix that included blurry, low-quality images, which is what gives it funkier, less sterile aesthetics than models trained only on clean stock.
- The open question. Where that lower-grade training data came from is not spelled out, and Addy called that unknown the thing that could block adoption in our industry.
Joey's argument: after being cool on LoRAs for a while, the specific looks coming out of this made the case for training your own again. Addy's counter: Krea's niche is precise style and feel, not universal capability like a Nano Banana, and the ComfyUI, human-in-the-loop setup keeps it feeling artist-driven rather than fully agentic.
What We Broke Down: LTX Trainer Brings Trainable LoRAs to Video-to-Video
LTX released LTX Trainer, a video LoRA of sorts. You teach the open LTX video model one specific video-to-video task, then apply it to your footage. The interesting part is that you train it on whatever narrow job you have.
Examples we walked through:
- Style transfer. A LUT-style look, a lens effect, a color treatment applied across a clip.
- HDR upsampling. Greg Tiergarten's experiment using the trainer for HDR upres, which we called thinking outside the box.
- A water model. Trained only on water, it adds rain, street reflections, and background rivers to dry footage, effectively a targeted simulation tool.
- Decompression. Feed it rough web footage and it comes back sharper, with deblocking and up-resing that helps technically without changing the style.
- A cross-eyed effect shown in a ComfyUI clip, plus the note that you can train audio alongside the video.
It felt like an open-source take on the "shoot footage, train on it, make more from it" idea we associate with Ben Affleck's InterPositive, which led to our running joke that maybe it was LTX under the hood the whole time. The golden use cases here will surface once people spend six months experimenting.
What We Questioned: Adobe Acquiring Topaz Labs, the Pipeline's End Node
Most AI video pipelines quietly end at Topaz Labs. It ran for roughly two to three decades as a non-AI upscaler built on GAN models, the go-to for pulling something usable out of old, rough footage, then pivoted into generative upscalers like Starlight and Astra. No matter what models sit upstream, Topaz tends to be the finishing node.
For a while, Adobe has shipped a Topaz integration, as a Photoshop plugin for upressing images and as a Firefly partner model, so folding it in fits the pipeline it is building. The deal keeps the team intact, with CEO Eric Yang still leading. Addy disclosed that Adobe is his day job and that his views here are his own, and Joey has interviewed the Topaz team for VP Land.
Addy's take: the current Adobe AI stack, Firefly Boards, Morph Cut, Nano Banana in Photoshop, and Harmonize, is small next to what he expects over the next couple of years, and he sees Topaz playing a key part in that rethink.
Bottom Line: Big-Studio Deals and Open Weights Are Meeting at the Workflow Layer
Everything in this episode pointed at the same layer: who controls model training, and where it happens.
- Google and A24. Google buys research reach into a respected studio, mostly to feed and test the world models that serve robotics, cars, and ads, with filmmaking tools as the visible part.
- Krea 2. Open weights and a training report put specific, stylized image models on local machines, with the training-data source as the open risk.
- LTX Trainer. Trainable video-to-video effects let one person build a narrow tool, from wet-footage simulation to decompression, without a giant lab.
- Adobe and Topaz. The pipeline's finishing step gets absorbed into a suite that wants to own more of the workflow.
Model training is spreading toward both ends at once: the biggest labs writing checks into studios, and open weights letting individuals train the same kinds of effects at home.
Links from This Episode
Companies
Google DeepMind, the investor and research partner in the A24 deal.
A24, the studio taking the research investment.
Adobe, acquiring Topaz Labs for its creative pipeline.
Midjourney, the image model now working in biomedical imaging.
Tools & Platforms
News & Analysis
WATCH THE FULL EPISODE ON YOUTUBE →
