On Denoised, we split “AI hybrid” filmmaking into two workflows, because the right way to shoot depends on whether any camera-original pixel survives to the final frame. We argue over green screen, gray box, and ordinary rooms. Then we cover a reported open-model acquisition, new Mac chips aimed at local AI, a claymation marketing spat, and two video models head to head.
Jump to:
00:42 A reported multibillion-dollar deal for the open-model hub
06:49 Green screen vs. normal rooms for AI hybrid shoots
25:19 New Mac chips pitched at local AI
34:17 Claymation marketing and a post-trained video model
46:00 Two new video models and their guardrails
Quick Take
Every backing decision on an AI hybrid shoot traces back to whether you plan to finish on camera footage or on model output. Control of the final pixels also shaped our takes on open models, local AI chips, and video-model guardrails.
What We Debated: Flat Color Voids May Hurt AI Hybrid Tracking, Depending on Your Pipeline
We took on Jon Finger's post arguing that the backings filmmakers reach for by habit work against AI hybrid shoots.
Finger, who has been running experiments at Luma AI, wrote that “White walls, green screens, grey walls, and other flat color ‘voids’ are worse for tracking your Ai Hybrid production than more natural environments.” As we discussed it, models trained on real-world footage have more to grab onto in a busy, ordinary room.
Two hybrids, two answers:
Gen AI-native hybrid. The photographed performance drives a synthetic result, with a model handling relighting, recostuming, and environment replacement before upscaling with tools like Topaz Labs.
Camera-first hybrid. You return to camera-original footage, with its full color data, and composite AI-derived elements through a traditional VFX process.
For native workflows, the busy room may win. A flat void gives a video-to-video model little to understand, and we've seen weird outputs from those plates even when they're covered in tracking dots.
We pointed to Beeble SwitchX demos of a South Korean period film shot in a cluttered office, with swapped backgrounds we thought looked great. As far as we know, depth models still train on real imagery, so a depth pass may separate a performer better in a normal room than in a gray void.
For camera-first workflows, backing is insurance. If the AI pass fails, a green or blue screen with trackers still gives you footage you can marker-track and composite conventionally.
Kino Flo MIMIKs take another route, combining green screen and projection.
Rotoscoping a cluttered background stays painful even with current AI tools.
Green and blue. Reliable for keying, but spill on performers remains the big cost. Neural keyers like Corridor Crew's CorridorKey target green screen's hardest problems.
Gray. Less spill, and AI roto tools select people well enough that a strong depth model might handle the matte, though we treat gray as an aid with no guaranteed one-click matte.
The gray box as gas lamp. One of us compared gray boxes with trackers to early fixtures pairing a gas flame with an electric plug for people who didn't trust electricity yet. That fits Hollywood VFX today, where a few shots get an AI attempt while teams build the traditional version anyway.
What We Questioned: Whether Performance-Driven Synthetic Pixels Will Outrun Camera-Original Finishing
One of us tested performance remapping with Google Omni and found camera placement mattered far less than expected.
The selfie test. A first attempt carefully matched camera position to the target character. A second used a phone selfie of the line reading, and Google Omni remapped it to a synthetic character at a new angle and pose while keeping the expression and delivery.
That was a personal, unbenchmarked experiment, and it left us asking whether the background needs removing at all.
The case for native. One of us argued that performance is the hardest element to get by prompting synthetic people, while lighting, costume, hair, and makeup can move to post. That view predicts a standard workflow, maybe a couple of years out, where performers shot with basic cameras in normal spaces drive synthetic versions of themselves.
The case for real actors. The other expects premium theatrical work, union protections, and 4K delivery specs to keep real actors on camera, with their performances preserved while environments change around them. He also sees fully synthetic casts already appearing in AI vertical dramas.
We landed on a freeway image, with SAG-governed, camera-pixel theatrical production as a big rig at 55 miles per hour and the native workflow as an electric car passing it. One of us predicted 10 to 15 films under $10 million in the faster lane for every large analog production.
What We Questioned: Why NVIDIA Would Reportedly Pay $12.9 Billion for an Open-Model Hub
We tried to work out what chipmaker NVIDIA gets from a reported agreement to buy the platform where creators download open models.
Reuters' account, relayed by Channel News Asia, puts the deal at $12.9 billion, citing The Information, and says neither company had confirmed it at publication.
The target is Hugging Face, where open models for ComfyUI workflows have long lived.
NVIDIA was already an investor in that distribution layer for open models spanning text, image, video, audio, and 3D.
Where's the revenue? Downloads are free with an account, and the business we could see is paid storage and subscriptions, similar to GitHub. We couldn't work out how that grows into a price that size.
Our read on the motive. We dismissed online theories that NVIDIA wants to shut open models down. One of us argued that local model users need NVIDIA cards, so a healthy open ecosystem hedges against closed labs like OpenAI walling off the frontier.
What We Explored: Apple's New Mac Chips Pitched at Local AI
We looked at Apple's Mac mini and Mac Studio update, which brings M6 and M5 Ultra chips without a design change. Much of the framing leans on local AI and always-on agents, the use case that sent people buying Mac minis to run OpenClaw.
The CUDA question. As far as we're aware, most models are built around NVIDIA's CUDA, and we weren't sure how well they run on Apple's neural accelerators, or whether a large model like Kimi feels the same as on NVIDIA hardware.
For editors, buy storage first. One of us runs Premiere on an M5 with 4K monitoring without hiccups. For editing rather than local models, a bigger internal drive beats a beefier chip, especially with the base Mac mini at $899 by our count, up from $599.
What Apple might be setting up:
A local model tuned for its own silicon, which would fit Apple's privacy branding.
An assistant that helps run complex apps like Resolve or Nuke, handled more gracefully than the Windows feature that screenshotted desktops every few seconds.
What We Debated: fal's Sub-Dollar Claymation Reply and Its H3 Max Model
Apple's Mac mini spot was claymation, and Apple posted a behind-the-scenes look showing it wasn't AI.
As one of us recounted, the team at fal promoted its own claymation reply as made in 15 seconds for 90 cents.
The split. One of us called it cheesy given the handwork behind Apple's piece, and the other saw an effective way to get attention. We agreed the bigger friction comes when quick AI work riffs on someone else's original idea or IP, like generated Batman or Superman clips.
Does process matter to audiences? We raised Matt Workman's point that most viewers already assume CGI is AI, and we added that many don't care. We think the million or so people who make what billions consume still haven't settled how to use these tools.
H3 Max. fal's launch post calls H3 Max a new post-trained video model from fal Research and claims improvements in quality, prompt understanding, aesthetics, and generation speed. As we understood it, fal post-trained a MiniMax model and offers the result through its own platform.
Training a house style. That raised a question about training a video model on your own animation style, the way one of us recalled someone at fal training one on the Fleischer look. The front end stays artist-driven, with look development and production design feeding a model instead of rigging and rendering.
What We Tested: Wan 3.0 Against Seedance 2.5's Likeness Guardrails
One of us ran the same self-uploaded input through Alibaba's Wan 3.0 and a competing model.
In ByteDance's Seedance 2.5, that input triggered likeness alerts and blocked generation.
Wan 3.0 produced output we put close to Seedance 2.0 quality, with practically no restrictions.
Seedance 2 let some uploads through, while 2.5 felt completely locked down in our use.
The counterpoint. Wan output still reads a bit AI-y next to Seedance 2.5 or MiniMax. What we understood as one of its pitches, assembling a 30-second trailer or commercial from PDFs and images, looks aimed at social content at scale, a fully automated approach we didn't warm to.
Open weights went away. Wan 2.2 became an open-weight standard, 2.5 was the first API-only release, and 3.0 is API-only too. We think open weights make the most sense when the parent sells cloud compute, as Alibaba does, and when the model fits a single consumer GPU.
We keep asking what the business case is for each open model, and the reported Hugging Face price sharpens that question.
Bottom Line: Control Is Shifting From the Camera to the Model Stack
Each segment came back to who controls the final pixels and the systems that make them.
Hybrid shoots. A normal room may serve native workflows, while green, blue, or gray remains the fallback when camera-original pixels reach the final frame.
Performance remapping. Our selfie test suggests performance may matter more than camera placement, though we expect theatrical specs and union protections to keep real-actor pipelines in place.
NVIDIA and Hugging Face. The reported deal would tie where creators get open models to the hardware they run on, with business logic that stays unclear to us.
Apple silicon. The new chips are framed for local AI, and CUDA dependence is the open question.
fal and claymation. A cheap generated reply drew attention, and H3 Max shows a hosting platform turning post-training into its own model.
Wan vs. Seedance. In our test, guardrails separated Wan 3.0 from Seedance 2.5 as much as quality did, and Wan has moved from open weights to API-only.
We think the people making most of what audiences watch are still deciding how to use these tools, and their choices will shape entertainment for the next few years.
Links from This Episode
Tools & Platforms
Luma AI: Where Jon Finger runs hybrid experiments.
Beeble SwitchX: Period-film demo shot in an ordinary office.
Google Omni: Our phone-selfie performance-remap test.
Topaz Labs: Upscaling for finishing AI-native output.
fal H3 Max: Post-trained video model from fal Research.
MiniMax video model: The base model we understood fal post-trained.
Wan 3.0: Alibaba's API-only video model we tested.
Seedance 2.5: ByteDance model that blocked our likeness test.
Companies
News & Analysis
WATCH THE FULL EPISODE ON YOUTUBE →
