CDance 2.5 makes AI video whole-person replacement shockingly convincing — and ByteDance isn't open-sourcing it

CDance 2.5 Is Shipping Full-Body Replacement—and ByteDance Isn't Open-Sourcing It

CDance 2.5, ByteDance's latest video synthesis model, replaces entire people in footage with photorealistic fidelity that's tricking viewers into mistaking AI recreations for real recordings.

The shift from face-swapping to full-body replacement marks a meaningful jump in capability. The model handles subtle motion consistently—occlusion (objects passing in front of the subject), variable framing, and motion blur all integrate without obvious artifacts. It works equally well whether the target person is small in the frame or large, looking at the camera or away.

The viral signal

SemiAnalysis founder Dylan Patel posted a video of himself as Drake in a cheetah-print music video over the weekend. Most viewers initially believed it was real. The video didn't saturate the internet until hours later, and it came out the day after Drake's actual music video dropped, making timing plausible enough to defeat initial skepticism. Patel stayed in character throughout multiple shots—walking around, entering a mirror, ascending stairs—without obvious facial drift or body distortion.

The deception worked partly because of the underlying footage's quality and the model's handling of motion, but also because viewers lacked prior exposure to the authentic video as a reference point. Once the reference existed, the gap between the AI version and real footage became apparent in subtle ways: facial structure occasionally reverted to Drake's features, and fine details occasionally degraded.

Mechanism and distribution

The production pipeline chains multiple tools together. Codex (an AI workspace and agent layer) runs on a local PC and connects via API to ChatGPT for reference image generation, then routes footage through CDance 2.5 via Higgs Field's API wrapper. The user controls everything from a phone via ChatGPT's app interface without writing code.

Access is neither open-sourced nor freely available. CDance 2.5 runs through Higgs Field's commercial API, which charges by consumption. The pricing model itself is problematic—users report confusion between plan credits and API credits that don't interchange, making it difficult to predict spending.

Why ByteDance keeps it closed

CDance remains proprietary rather than open-sourced, which signals strategic intent. The model is compute-intensive to run. China faces meaningful constraints on GPU access relative to American labs, so exporting the capability would increase American researchers' ability to iterate faster while consuming Chinese compute. Keeping the model gated behind an API lets ByteDance monetize the capability while controlling distribution and preventing domestic rivals from weaponizing it.

The technical bar is high enough that open-sourcing would be meaningful. Unlike earlier face-swap tools (DreamBooth), which hobbyist engineers rapidly ported to consumer GPUs, CDance 2.5 resists commodification. The computational requirements stay prohibitively high for local inference.

Near-term use cases

Hollywood is already adopting these tools for reshoots. The Henry Cavill Superman example illustrates the previous state: removing a mustache in 2017 via manual VFX reconstruction was tedious, expensive, and visibly poor. CDance 2.5 handles the same task in seconds with superior results.

Most fictional performances will likely continue being shot traditionally—actors perform, and the footage remains untouched. Reshoots, stunt doubling, and scenes where the actor is unavailable represent the immediate wedge. The tool accelerates production where it solves a concrete problem (actor availability, continuity) rather than replacing the creative act of performance itself.

The commodification cycle

The trajectory mirrors DreamBooth's arc: impressive demo → everyone makes their version → becomes meme → melts into production background. What feels novel for a few weeks (replacing yourself in iconic videos) will become routine. The technology stops being the story; it becomes the baseline for how footage gets made.

The real friction point is access and pricing, not capability. Higgs Field's confusing billing structure (multiple credit types, plan/API separation) is already deterring users despite the underlying product being genuinely useful. A cleaner API wrapper or competitor entry could accelerate adoption.

The open question ByteDance is avoiding

The model works. Watermarking and provenance claims are absent from the transcript entirely. If synthetic footage becomes visually indistinguishable from real footage at scale, the downstream problems—deepfakes, identity fraud, legal liability for platforms hosting the output—arrive whether the technology is proprietary or open. ByteDance's strategy buys time and revenue but doesn't solve the governance problem.