Summary
Nick Saraev demonstrates a full AI-powered video style-transfer pipeline, arguing that the ability to apply custom visual styles to recorded footage represents a fundamental shift for content creators — comparable, he suggests, to how fashion or jewelry once expressed identity. The entire video is delivered in an animated style as a live proof of concept.
The technical pipeline works as follows: a source video is recorded against a clean background, a style reference image is generated using GPT Image 2.5, and then the Genjutsu model on Higgsfield applies motion-preserving style transfer — combining the reference’s aesthetic with the source video’s head movements, lip articulation, and body motion. Because current models cannot process a full video in one pass, GPT-6 Astra acts as production manager, breaking footage into 20-second chunks, inspecting every join for continuity errors, retrying failed generations, and stitching outputs together. Post-processing uses Solero for voice activity detection and ffmpeg for frame-rate control.
Saraev covers two paths: a consumer-accessible API method and a scalable production pipeline optimized for cost using smaller models like One Three Turbo for non-critical steps. He notes the technique is particularly powerful for localized advertising — one source video, many stylistically varied outputs — and shares that all templates and prompts will be open-sourced below the video. The main current limitation is per-clip length constraints on Genjutsu, which he expects to improve within three to six months.
📺 Source: Nick Saraev · Published September 11, 2026
🏷️ Format: Tutorial Demo
