CorridorKey 2.0 is out, and its keyer now reads the whole clip instead of single frames to pull cleaner green screen keys. On Denoised, we break down how Corridor Crew retrained it, weigh the new Kling 4.0 model's per-second pricing, and question when native 4K generation is worth paying for.
Jump to:
- 00:17 CorridorKey 2.0 and temporal keying
- 07:38 Detailed 3D worlds built entirely in code
- 16:25 Kling 4.0 features and per-second pricing
- 27:15 Ideogram's vector model and FLUX 3 Max
- 30:12 Native 4K generation vs. upscaling from 720p
Quick Take
Every upgrade in this episode carries a cost question. CorridorKey 2.0 is a one-time $50, Kling 4.0 runs about half the per-second rate of ByteDance's Seedance 2.0 on the same platform, and our guess for one Seedance 2.5 4K clip is $20 to $30. We sort out which upgrades change a working pipeline and which mostly burn budget.
What We Explored: CorridorKey 2.0 Reads Whole Clips for Cleaner Green Screen Keys
Corridor Crew's CorridorKey keyer pulls a key from green screen or blue screen footage. We've had good results from the first version with some tweaking, and 2.0 goes after the shots that usually break keyers.
The first release was an open-source neural network you had to know how to run yourself, with community-built plugins around it. Version 2.0 keeps the weights free and open source.
Temporal awareness is the core change. In our summary of Corridor's training breakdown, version 1.0 was trained frame by frame, while 2.0 looks across the entire clip, backward and forward in time.
Corridor brought in neural network specialists and trained the model on local Puget Systems hardware instead of a render farm or cloud compute.
The demo shots target the hardest keys. Our favorite shows a subject on the green going out of focus, another features a glass of water, and the demo pulls clean keys from both.
Addy's test: He bought the $50 license and ran a 150-frame 4K clip in about a minute on what he believes is an RTX 4090. He suspects it's optimized for CUDA and favors NVIDIA GPUs, though a Mac beta is also available.
What the $50 one-time fee covers:
- Standalone app. Key outside an editor entirely.
- Host plugins. Resolve, Premiere and most other major apps.
- A friendlier interface. No need to set up the open model yourself.
Addy's take: Next to roughly $300 for Blackmagic Design's Resolve Studio, $50 is an easy call. Joey's argument: He'd pay it to support the project even if he planned to run the free weights.
Early reactions are positive. VFX artist comments Addy has seen are mostly favorable, and we think 2.0 does especially well on movement, messy hair, skin highlights and despill.
Addy's speculation on training: A strong depth model under the hood could explain clean keys across dark, light and patchy green. Paired data, raw plates matched with artist-corrected composites, would carry compositor judgment into the model.
Our recollection is that Corridor described synthetic data for the first release, built from MetaHumans on green paired with finished keys.
Joey's tip: On a plate that wasn't evenly lit, he had Claude run two passes, one cranked up to hold hair detail and one dialed back for a cleaner key at the feet. Describing the goal and sharing visuals is enough for an agent to work out that kind of workflow.
Joey's argument: 2.0 handles green and blue, and his prediction for 3.0 is gray screen or no screen, since he sees gray as the go-to backdrop for AI hybrid productions. The case for green or blue, from our green screen debate, is that a keyer like CorridorKey and traditional compositing give you a fallback when AI tools struggle with gray.
What We Questioned: Leon Lin's Detailed 3D Worlds, Built Entirely in Code
Leon Lin has posted elaborate 3D worlds made entirely in code and rendered in Three.js, the web library for drawing 3D in a browser, with no Blender control involved.
In the repos and prompt he shared, he generated nine reference images and asked for "a massive ultra-detail 3D world in Three.js based on the 9 reference images I attached," with the rule "Zero external assets. Everything has to be generated in code."
Addy's take: The quality already matches a 2000s video game, and we both thought of Myst. It ran on a laptop GPU, a Radeon 780M, and he wonders what an RTX 5090 could do.
A second scene swaps in Python and GLSL. Lin's Alpine lake world uses GLSL, a graphics shading language that writes shaders as math.
Addy's team built a GLSL fur shader for work around 2012, and he described how hard it was:
We had to bring in a PhD who understood GLSL and mathematics at that level to be able to write something that mimicked fur, because fur is so heavy, it uses geometry. So this GLSL shader mimicked fur without any of the geometry.
That shader layered 2D cards angled to the camera. In Addy's words, Claude now runs that class of work "on autopilot," and he stressed that "there is nothing generative here," calling it "purely an LLM coding."
The Alpine scene cost a little under $3,000 at API prices, per the numbers Lin posted for his Claude usage. Our guess is that he used a subscription such as Max and reported the API-rate equivalent, since subscriptions are heavily subsidized.
Joey's argument: Lower-fidelity versions of this already power browser-based 3D blocking tools, where you block a shot and feed it to a video model for camera control. This level of detail reads more like an experiment, and we're curious whether it could run in real time as an explorable world.
What We Debated: Kling 4.0 Matches Seedance 2 Feature-Wise for About Half the Per-Second Price
Kling 4.0 is out after a quiet stretch following Kling 3, while Seedance 2 drew most of the attention. From Kling's 4.0 release notes, the model is set to support:
- Output up to 4K, including 4K 10-bit HDR
- A 21:9 cinematic aspect ratio
- Two-channel stereo output, which we weren't sure how to read
- Up to 10 image inputs
- Up to five video inputs with a combined duration of up to 30 seconds
- Audio input for voice reference on speaking characters
We don't think all of these are out yet, at least not on fal's model platform.
Joey's argument: Feature-wise, this is Kling catching up to Seedance 2, and pricing is what stands out. He'd also like Kling's separate motion transfer model, which mapped a human performance onto an AI character, to return in some form.
Per-second rates on fal at 720p with audio:
- Kling 4.0. About $0.15 per second.
- Seedance 2.0. About $0.30 per second.
- Seedance 2.5. About $0.47 per second.
That puts a 10-second Kling 4.0 clip at about half the cost of Seedance 2 and a third of Seedance 2.5. Seedance rates on fal run high compared with other providers, so treat these as a baseline.
How is it this cheap? Addy asked whether Kling is losing money; Joey wondered whether it found a way to deliver the quality with less compute. Addy noted Huawei is reportedly building data center processors to rival NVIDIA's, which could lower costs for Chinese AI companies.
The sample clips are mixed. Some still look synthetic to us, close-ups hold up, and wide shots are where things fall apart. Handheld pans show more natural motion blur than Kling 3's, and a 15-second action sequence plays as a single shot.
Addy's take: AI faces are getting closer to real than Unreal Engine MetaHumans ever did, since those stayed in the uncanny valley.
What We Explored: Ideogram's Native Vectors, fal's 10-Second FLUX 3 Max and Leaner Architectures
Ideogram released its first vector model, which Addy reads as native vector generation rather than a pixel-to-vector conversion. His own pixel-to-vector attempts turned out poorly, so native vectors could fill that gap for graphic design work.
fal post-trained FLUX 3 into FLUX 3 Max. By fal's speed chart, generation drops from about two minutes to about 10 seconds, and fal's own ranking puts the quality higher.
Addy's question: Why use the original if fal's version wins on both counts? Joey's answer: The open weights still run on your own hardware, and he assumes, without being certain, that Black Forest Labs takes a licensing cut when its models run on fal.
Next-concept modeling targets AI's compute bill. Addy flagged it as a non-transformer language model architecture, with figures claiming text generation 10 to 31 times faster and four to five times fewer training tokens.
Jev trades sophistication for speed. Addy described it as a simple gating model; Joey added that its advantage is speed and cost on fuzzy, unstructured data.
What We Questioned: When Seedance 2.5's Native 4K Beats Upscaling From 720p
ByteDance is rolling out Seedance 2.5 4K with no pricing posted. Seedance 2.5 at 1080p on fal runs about $7 per 10 to 15 second generation, so our guess for 4K is $20 to $30 per generation.
Joey's argument: Work at 720p until the shot is right, then move to 4K with everything locked. We've had decent results feeding a depth map of the approved 720p clip, plus the original reference images, into the 4K generation, which keeps the output consistent depending on the shot.
Seed locking hasn't worked for us, since the same seed and inputs at 4K still produced a different clip.
Addy's counter: Each 720p pixel now has to fill many 4K pixels, which makes it a different problem for the model. Upscalers make sense as a finishing pass, and paying 4K rates to experiment gets expensive.
Where native 4K earns its cost. Plenty of creators generate strictly at 720p and finish with Topaz Labs' upscaler or similar. Wide, detailed shots are still AI's weakest area, and at 720p their small details give an upscaler too little to work with.
- Detailed wides favor native 4K. Crowds, buildings full of lines, or a 1970s New York street with people and taxi cabs should come out better when 4K references feed a 4K generation.
- Close-ups don't. For faces, we see no benefit to generating in 4K, only extra spend.
Addy's framing is that native 4K is a lever to pull when the shot calls for it.
Bottom Line: Pay for Quality Where the Shot Needs It
Across the episode, every better result came with a clear cost, and the useful skill is knowing where to spend it.
- CorridorKey 2.0. Temporal keying handles hair, motion and despill better in our view, with free open weights and a $50 app.
- Code-built 3D worlds. Agents now handle code-heavy techniques that once needed specialists, at a little under $3,000 in API-equivalent usage for one scene.
- Kling 4.0. Feature parity with Seedance 2 at about half its per-second rate on fal.
- Ideogram, FLUX 3 Max and leaner architectures. Native vectors, faster post-trained models and non-transformer designs all aim to cut time and compute.
- Seedance 2.5 4K. Worth it for detailed wide shots, wasted on close-ups.
We expect top models to get cheaper as newer ones arrive, so iterating at low resolution and finishing high stays useful as prices shift.
Links from This Episode
Tools & Platforms
Projects
Companies
Research & Reports
News & Analysis
WATCH THE FULL EPISODE ON YOUTUBE →
