In this episode of Denoised, we work through a batch of incremental model updates and one licensing change from Epic. We test Pika's open-weight H3 Flex, compare how Google's Omni and ByteDance's Seedance treat cinema-camera footage, and debate Epic letting AI models train on Quixel Megascans alongside a GPT workflow that turns phone scans into CAD files.
Jump to:
00:39 GPT Image 2.6 adds GIFs and reference images
05:38 Omni 1.2 Flash and how it handles real camera footage
11:06 Pika H3 Flex goes open weight
19:02 Quixel Megascans goes free and opens to AI training
24:39 Image to CAD with GPT
Quick Take
Most of the model news here, from GPT Image 2.6, the latest update from OpenAI, to Google's Omni 1.2 Flash, is gradual, with faster generations, longer shots, and per-second pricing. The bigger questions are how much of your original footage survives a video model, who controls the weights you build on, and what happens to entry-level technical roles when a model does hours of specialist work in minutes.
What We Explored: GPT Image 2.6 Adds GIFs and a Faster Flare Mode
GPT Image 2.6 reads more like a feature expansion than a major model update. The main addition is GIF output, plus saved reference images in ChatGPT so you can call up your own characters or pets.
- Two flavors remain. Sunburst is the heavier option for pro-level image editing, and Flare is the faster mode for ideation and volume.
- Flare got faster. OpenAI's comparison shows Flare in 2.6 generating faster than in 2.5.
- Saved references live in the app. That feature looks most useful if you work inside ChatGPT.
Annotation still does the heavy lifting. For edits, one of us still reaches for Nano Banana Pro 2 or GPT Image 2.5, because both read red-marked annotations well.
That workflow means redlining an image in Photoshop, feeding it in, and often skipping the text prompt or asking the model to follow the notes in the image. It works well for rearranging a frame and repositioning objects in 3D space on a 2D image.
Speed is the next behavior shift. With quality now usable for many tasks, waiting on a render is the friction that remains, since a slow generation pushes you to switch tabs and lose focus.
We compared it to Unreal Engine and Unity reaching real time on GPUs in the 2010s, which helped open up the LED volume industry and location-based AR and VR experiences. Real-time AI generation could set off a similar wave of tooling built on top of it.
What We Questioned: Omni 1.2 Flash Gets Longer Shots, but Inputs Still Limit It
Google's Omni 1.2 Flash rolls out with quality improvements and continuous shots up to 10 seconds, with 11 to 15 seconds available for Google AI subscribers. It also moves to per-second pricing, so you pay for the duration you generate, edit, extend, or upscale.
Seedance pricing is harder to predict. It is token-based and shifts with pixel count, reference video, audio, and other input and output factors. Per-second pricing ties Omni's cost directly to the duration you generate.
Inputs are the bigger gap. Omni handles video editing well, and in our read it accepts an input video and a start frame but not reference images. As far as we can tell, you can't hand it a video of a person plus an image of who that person should become, an input that is standard in Kling's video models and Seedance.
When we tested the original Omni, it turned a dog into a robot dog convincingly, but only through a text prompt.
Omni may keep more of your footage. One of us has found that Omni keeps film grain, noise, and texture from high-resolution cinema footage, like 6K from an ARRI Alexa, better than Seedance, which tends to mush the image because it generates every pixel.
The other of us hasn't tried that yet and wants to, since compressed-looking output from footage shot on good cameras is a recurring problem.
Models still try too hard.
- Seedance edit mode, even with a mask, can hallucinate or add things that were not in the footage. One remove-a-person request added a stranger in a black trench coat with an umbrella.
- A removed person can come back as a light pole that then turns into a dog.
- Kling 3 kept flying the camera around when the shot called for a locked-off frame.
One idea we want to try is marking up the keyframe with a reference image as a picture-in-picture inset, borrowing the mark-up-your-frame approach Veo kicked off. Omni appears to be a different model from Veo, so whether it reads those notes is an open question.
What We Tested: Pika H3 Flex Releases Open Weights Built on MiniMax
Pika's H3 Flex is a post-trained version of MiniMax, and Pika released the weights so you can download it and run it locally. Pika says it is faster than other MiniMax models, with better quality and prompt adherence.
We also tried the version on fal, and it was also very good. Both beat MiniMax out of the box on quality and adherence.
Post-training here means a LoRA. The open release is a post-trained LoRA that sits on top of the MiniMax base model. MiniMax's own H3 model unifies text, image, video, and audio generation, with open weights listed as coming.
What post-training takes.
- A goal, such as resolution, speed, a specific look, or camera motion.
- A dataset of images and video gathered around that goal, with captions.
- Epochs of training on the weights, checking progress along the way and correcting the strategy when it drifts.
- Serious GPU time. With a couple of H100s, a run could take a few days, which is why hosted services that train on your uploaded video, images, and captions exist.
What We Questioned: Steve Aoki's Archon Promises Raw Footage to a Finished Edit
In its post on X, Selendis describes Archon as a video foundation model that zero-shots raw footage into a finished edit in minutes. Steve Aoki is listed as a co-founder of Archon, and the demos lean on vertical video, which points to social content, a space where his music shows up frequently.
Clipping is harder than it looks. Our own show clips run through a Claude agent with instructions and visual checks, and it still regularly frames the wrong face, showing one of us paused mid-reaction over the other's audio.
Opus Clip handles speaker framing well because a dedicated team trained models on that one task. That gap raises the maintenance question that comes with any vibe-coded tool, which is who keeps it working. We plan to try Archon for clips and highlights.
What We Debated: Epic Makes Megascans Free Again and Opens Them to AI Training
Epic is making the full Quixel Megascans library free again, and the updated terms let you train AI models on Megascans and MegaPlants.
According to Quixel's announcement, "your AI models can now learn from real-world scans to help you generate varied photorealistic environments for your projects."
Megascans were a photorealism shortcut before Fab. Quixel was known for some of the best scans and fully built natural environments, much of it European forests and ruins. Before AI, packing scenes with Quixel assets was one of the main routes to photorealism, and after acquiring Quixel, Epic made much of the library free inside its marketplace.
- LED volume work leaned on it heavily, and one of us built an LED volume short almost entirely from kitbashed Quixel environments.
- Epic then acquired Sketchfab and merged its marketplaces into Fab, and the licensing changes put much of the library behind a paywall.
- Assets got harder to pull into projects, and that friction pushed one of us to use Unreal less.
One of us also saw a related report that Epic is shutting down the MetaHuman browser configurator, which would move that configuration work into the engine itself.
Our read on the AI terms. A library this large, full of texture maps and geometry, is exactly what a company training a model would want. There is probably little value left in paywalling it and more in giving it away so artists and model builders stay in the Unreal and Fortnite ecosystem.
What We Debated: GPT Turns Phone Scans Into CAD Models and Printable STL Files
In his scan-to-CAD video, Michael Lydick follows his hobby of scanning objects and reverse engineering them into CAD models. His phone scan is rough and serves only as a reference object, and instead of modeling around it himself, he had GPT do the CAD work, which he says compressed hours of modeling into minutes.
Why CAD is slow by hand. The work goes curve by curve and layer by layer, placing every fillet and spline, and it is parametric. Dimensions tie to each other through equations and dependencies, so resizing a box resizes the window with it.
Skipping CAD entirely. Lydick also says the model can go from a photo straight to an STL, the compiled file a 3D printer reads, bypassing the editable CAD file. One of us has done something similar for editing, going from a chat conversation straight to an EDL or XML without opening a timeline.
The case for caution. If you dislike the STL you get, you are stuck, because you need a CAD file to change anything. The deeper worry is the talent pipeline, since CAD fluency takes years, like learning a language by living in the country, and entry-level work is what a model like this absorbs first.
The same concern applies to After Effects and Nuke artists, who often spend a decade on smaller shows before landing a tier-one, $100 million film.
The case for access. For people with no 3D background, this changes 3D printing. Ideas for camera accessories that once stalled at the design step now look viable, with Claude modeling the part and a printing service producing it.
That helps printer makers like Bambu Labs, even if it leaves open what the next generation of CAD engineers looks like. We compared it to old photos of drafting rooms full of engineers with slide rules at drawing boards. CAD software collapsed that room into one computer, and in our view AI is compressing it again faster than the industry can respond.
Bottom Line: Control Over Inputs, Weights, and Skills
Every story in this episode comes back to how much control you keep when a model does the work.
- GPT Image 2.6. GIFs and faster Flare generations help, but annotation-driven editing remains the workflow that matters.
- Omni 1.2 Flash. Per-second pricing and what one of us has seen as better grain retention with Omni make it worth trying with real camera footage, though it needs reference-image inputs.
- Pika H3 Flex. Open weights for a fast-iteration video model let you run and build on it locally, with output up to 4K.
- Archon. Automated editing for social is promising, and accurate speaker framing is the test it has to pass.
- Quixel Megascans. Free assets with AI training rights could keep artists and model builders inside Epic's ecosystem.
- Image to CAD. Faster 3D design opens printing to non-specialists while making it harder to see how entry-level technical artists will learn the craft.
Across all six stories, control sits in reference inputs, open weights and licensing terms, and the CAD and compositing skills needed to fix model output.
Links from This Episode
Tools & Platforms
GPT Image 2.6, OpenAI's image model update adding GIFs and reference images
Omni 1.2 Flash, Google's video model update with longer shots and per-second pricing
Pika H3 Flex, Pika's open-weight, post-trained MiniMax video model
Archon by Selendis, an AI video editing model, with Steve Aoki listed as a co-founder
Quixel Megascans, Epic's scan library, free again with AI training terms
Projects
WATCH THE FULL EPISODE ON YOUTUBE →
