TeleStyle brings reference-based style transfer to video

Turning a photo into a watercolor painting is old news for generative AI. But trying that on video usually ends in a mess: either the characters warp into unrecognizable blobs, or the whole clip flickers so aggressively it becomes unwatchable.

Researchers introduced TeleStyle, an open-source style transfer model designed to tackle that exact headache. Built to handle both still images and moving video, it can restyle footage into comic books, Victorian paintings, atomic-punk illustrations, or minimal sketches while locking down the original content.

Most modern Diffusion Transformers struggle with this because their internal layers tangle visual content and style tokens together. Tweak the artistic vibe, and the subject collapses. TeleStyle works around this by building directly on Qwen-Image-Edit, inheriting a base model that already excels at keeping subjects intact during edits.

Smarter training over brute force

Training a model to understand every niche artistic style is usually a data nightmare. Curated datasets are too small, while synthetic web scrapes are noisy and erratic.

The team tackled this with a curriculum continual learning setup. They gathered clean, curated datasets of distinct aesthetics alongside thousands of synthesized triplets across messy in-the-wild styles. By gradually feeding this hybrid data to the model, TeleStyle learned to generalize to unseen art styles without wiping out the original subject matter.

For motion, the researchers added a dedicated video-to-video stylization module. Instead of running isolated image passes frame-by-frame, it enforces temporal consistency across the timeline. You can steer the aesthetic via text prompts or just hand it a reference picture, and the resulting footage keeps its structural integrity.

According to the report, the model achieved state-of-the-art results across three benchmarks: style similarity, content consistency, and overall visual quality. The weights and source code are already open to the public.

Why it matters

Video stylization has been stuck in an awkward middle ground. Commercial tools are locked behind subscriptions, and raw open-source pipelines are notoriously jittery. TeleStyle gives solo animators, video editors, and indie game developers a viable tool to completely re-theme footage using nothing more than a mood board image.

We think the smart move here was skipping the temptation to train a foundation model from scratch. Building atop Qwen-Image-Edit gave the team rock-solid content preservation out of the box, allowing them to focus entirely on disentangling style and smoothing frame transitions.

Keep in mind that while the technical report calls TeleStyle lightweight, running full video-to-video diffusion still demands serious GPU muscle, especially for longer high-res clips. But as an open-source release, it lowers the barrier for creative workflows that used to require a dedicated VFX department.