Summary
Creator Youri van Hofwegen walks through a complete workflow for maintaining consistent AI characters across multiple video scenes using Higgsfield, an all-in-one AI video platform. The video opens by demonstrating the core problem: generating the same character three times with identical detailed text prompts in Sedance 2.0 produces three visibly different faces, proving that written descriptions alone cannot lock down a character’s identity.
The solution is a multi-panel character sheet — a single reference image generated with GPT Image 2 at 4K resolution on a neutral gray background (chosen deliberately to avoid the exposure bias introduced by white or black backgrounds). The sheet contains three panels: a headless full-body front shot to establish outfit and proportions, a rear shot with the head included, and a tight facial closeup. This reference is then saved as a named element in Higgsfield’s element library, allowing it to be tagged in any future prompt without describing the character at all.
Van Hofwegen stress-tests the workflow across five scenes for a character named Elias — a rocky mountain ridge, a tent in a blizzard, a snowbound cabin, a dim tavern with a second character, and a map shop — using Sedance 2.5 (noted as a significant upgrade from 2.0 in consistency and motion quality). In every scene, including dialogue clips with lipsync, Elias holds his face, outfit, and proportions. The video also covers outfit and art-style swapping without losing facial identity. Settings like 15-second clips at 1080p 16×9 are specified throughout, making the workflow directly reproducible.
📺 Source: Youri van Hofwegen · Published August 13, 2026
🏷️ Format: Tutorial Demo
