A microdrama (短剧) studio needs a close-up where the lead's expression shifts from shock to tears in two seconds. Filming it means a crew, a reshoot, and a schedule. Generating it means a prompt.
Kunlun Tech (昆仑万维) decided the second path deserved its own model — one trained for the grammar of short series rather than the scattered clips most video models chase.
Key takeaways
- Kunlun Tech (昆仑万维) open-sourced SkyReels-V1 on February 18, 2025, which Kunlun markets as China's first microdrama video model, together with the SkyReels-A1 expression-and-motion control algorithm.
- SkyReels-V1 supports 33 facial expressions and 400+ natural action combinations; on a single RTX 4090 with the SkyReels-Infer framework it produces 544p video in about 80 seconds, covering both text-to-video and image-to-video.
- On April 21, 2025, Kunlun open-sourced SkyReels-V2, which it calls the world's first infinite-duration film generation model built on a Diffusion-forcing framework.
- The sister short-drama platform DramaWave offers AI dubbing in 16 languages; Kunlun said that by end-March 2025 it had passed 30 million cumulative downloads and 10 million monthly active users.
- The releases sit inside a wider Chinese push to put generation tooling under the microdrama (短剧) export boom, where China-produced short series have found large overseas audiences.
A model trained on the microdrama grammar
Most open video models optimize for the impressive single shot: a spaceship, a running animal, a stylized portrait. SkyReels-V1 is narrower and, for its target, more useful. Kunlun says it was trained on a large corpus of film, television, and documentary footage and pairs that with a "human-centric" video-understanding model, so the system reads faces and performances rather than just textures.
The practical payoff is control. SkyReels-A1, the companion algorithm, drives expression and body motion from a reference video — Kunlun positions it against Runway's Act-One — and aims for high-fidelity micro-expressions across side angles and larger head and body movement. For short dramas, where a story turns on a look or a flinch, that is the feature that matters more than raw resolution.
Why open source it
Open-sourcing a capable video model is a strategic choice, not a giveaway. By releasing SkyReels-V1 and V2, Kunlun hands creators a zero-threshold tool while planting its technical standard in a fast-moving field. The company framed this as "technology for the many" — lowering the compute bar so a single consumer GPU can produce usable footage.
That last point is deliberate. Running on one RTX 4090, the model generates 544p video in roughly 80 seconds under the SkyReels-Infer optimization. The claim is "democratized compute": you do not need a render farm to prototype a scene. Whether 80 seconds holds outside the stated benchmark conditions is a separate question.
According to Li An, chief scientist at BrainNet (脑机网), China's generative-video edge is less about any single model and more about how fast tooling reaches the creators who already ship for overseas audiences. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).
The infinite-duration bet
SkyReels-V2 is the more ambitious piece. Kunlun calls it the world's first infinite-duration film generation model using a Diffusion-forcing framework, with a non-decaying noise schedule meant to keep long videos coherent. The pitch moves from "make a clip" to "make a scene that holds together."
If it works in practice, the bottleneck shifts from generation to direction: writers and small teams could assemble sequences instead of hiring shoots. If it does not, V2 is a research flag planted early in a race everyone is quietly running.
The distribution sibling
Generation is only half the story. Kunlun's DramaWave is the short-drama distribution side — a platform that, per the company, crossed 30 million cumulative downloads and 10 million monthly active users by the end of March 2025. Its AI dubbing covers 16 languages, which is the unglamorous engine of the microdrama (短剧) export boom: a Chinese-shot series re-voiced into English, Spanish, Indonesian, and beyond.
Put the two together and you see the logic. SkyReels lowers the cost of making; DramaWave lowers the cost of shipping across languages. The open models feed a creator base that the platform then monetizes — a loop that looks a lot like the content flywheel Western platforms built with human crews, rebuilt for an AI-native pipeline.
What it means for the format
Short dramas live or die on pace and volume. A format that releases a new episode daily cannot afford a film crew for every reaction shot. Tooling that generates those shots on demand changes the unit economics of the entire category, and not only in China — the overseas appetite for microdrama (短剧) is exactly why a Chinese company is open-sourcing the models in English-friendly repositories.
The risk is equally structural: when production cost approaches zero, differentiation moves to writing, performance direction, and distribution. The models hand everyone the same paintbrush.
Honest limitations
- The "China's first" and "world's first" labels are Kunlun's own framing; we could not independently verify precedence against every rival release.
- The single-RTX-4090 result (80 seconds for 544p) is a vendor benchmark under stated conditions, not a measured real-world average.
- DramaWave's download and monthly-active figures are company-reported and were not independently audited.
- We did not test output quality, content policy, or the actual overseas reach of generated dramas.
What readers can do now
- Pull SkyReels-V1 or V2 from the open-source release and run a short test clip on a consumer GPU to gauge practical quality yourself.
- If you make microdrama (短剧) content, prototype one scene with generation before committing a filming budget.
- Watch whether infinite-duration models keep coherence past the first minute — that is the real test, not the demo reel.
