New ComfyUI Node Enables High-Quality Video Stylization on 6GB VRAM Cards

Revolution in AI Video Styling: TeleStyle Node Breaks VRAM Barriers

A groundbreaking development in the generative AI community is reshaping how creators apply artistic styles to video content. A new custom node for ComfyUI, developed by an anonymous AI engineer under the username hackerzcity, successfully implements the TeleStyle algorithm with unprecedented efficiency—enabling high-fidelity image and video stylization on graphics cards with as little as 6GB of VRAM. Previously, such tasks demanded high-end hardware with 16GB or more, excluding most hobbyists and indie creators from professional-grade results.

The innovation centers on solving a long-standing flaw in AI-driven video style transfer: temporal inconsistency, colloquially known as the "zombie morphing" effect. This phenomenon manifests as unnatural, flickering distortions across frames, where facial features and objects appear to warp and reconfigure unnaturally from one frame to the next. Standard pipelines, including those based on Wan 2.1, struggled to maintain visual coherence over time, rendering outputs unusable for professional or even amateur video projects.

The developer’s solution is elegantly simple yet profoundly effective. By treating the reference style image as "Frame 0" of the video timeline, the node establishes a stable visual anchor. The process begins by extracting the first frame of the input video, applying the desired style to that single frame, and then using that stylized frame as the foundational reference for subsequent frames. This ensures that the model "pushes" the style forward consistently, eliminating the chaotic morphing seen in other implementations. A comparison video posted by the developer demonstrates a stark contrast: where conventional methods produce disorienting, ghost-like artifacts, the TeleStyle node delivers fluid, painterly transitions that preserve motion integrity.

Performance enhancements further amplify accessibility. The node introduces a TensorFloat-32 (TF32) toggle, leveraging NVIDIA’s precision-optimized computing format. On RTX-series GPUs, this reduces generation time from approximately 3.5 minutes per clip to just under one minute—a nearly 70% speed improvement without sacrificing output quality. For users with severely constrained hardware, the developer offers a clever workaround: the diffsynth_Qwen-Image-Edit-2509-telestyle model can be loaded as a lightweight LoRA adapter within a standard Qwen workflow. This approach consumes a fraction of memory while retaining 95% of the original style transfer fidelity, making TeleStyle viable even on entry-level gaming GPUs.

The release includes a fully documented workflow JSON and an open-source GitHub repository, inviting collaboration and refinement from the broader ComfyUI community. According to early adopters, the node has already been integrated into indie film pipelines, social media content studios, and educational AI courses. Its low hardware barrier is particularly significant in regions where access to high-end computing is limited, democratizing access to cinematic visual effects once reserved for Hollywood VFX teams.

As AI-generated media continues to blur the lines between photography, painting, and cinematography, tools like this TeleStyle node represent a critical inflection point. They shift the paradigm from exclusive, resource-intensive systems to inclusive, community-driven innovation. With its elegant fix to a persistent technical flaw and its commitment to accessibility, this development may well be remembered as the moment video stylization moved from experimental curiosity to practical artistry.

Resources: GitHub Repository | Workflow Guide | Comparison Video