My AI drew the hero as a different man in every scene. It's one guy's life story
I build explainer videos fully automatically. For each scene, a local image-generation AI draws that moment, one frame at a time. I was making one about an ancient Roman general — the whole life of a single man. When it was done, I watched it back from the top, scene by scene. And then I froze. The hero is a different person in every scene. In one shot, a king wearing a crown. In the next, a…
An AI-generated animation depicting the life of a single ancient Roman general exhibited a notable inconsistency - the hero's appearance varied considerably across different scenes. In one scene, the hero was a king wearing a crown, while in the next he was a young soldier in a plumed helmet, and then an old man with a white beard. This inconsistency raised questions about the reliability of the AI's storytelling capabilities.
The root cause of this issue was the independent generation of each scene. The AI was prompted separately for every scene, resulting in the hero's face and historical setting being reimagined from scratch each time. This led to a "drifting" effect, where the shared characteristics of the hero's appearance and historical context were not consistently maintained.
To address this issue, the reporter implemented two key strategies. Firstly, they introduced a "character card," a single source of truth containing the hero's appearance details, such as age, hair and beard, armor, and cloak. This card was referenced in each scene's prompt, ensuring that the hero's appearance remained consistent across all scenes. Additionally, shared historical constraints, such as depicting the hero in the ancient era without a crown or dome, were also declared once and injected into every prompt.
However, despite these efforts, some inconsistencies still persisted. The reporter discovered that negative prompts, such as "don't draw a dome" or "don't draw text," were not entirely effective in preventing the AI from generating anachronistic elements or fake historical elements. To further improve the quality control, the reporter implemented an additional checkpoint that inspected the generated images themselves, rather than solely relying on the prompts.
This checkpoint verified that the generated images accurately depicted the intended scenes, checked for the presence of the character's same face, and ensured that the historical context matched the narrative.
By combining these strategies - injecting declarations into the prompts and implementing a pixel-level visual checkpoint - the reporter successfully achieved a consistent portrayal of the hero throughout the animation. The hero's face remained unchanged, and the historical setting accurately reflected the ancient era. This approach eliminated the need for manual visual quality assurance, automating the process and ensuring that the final product adhered to the intended narrative and visual standards.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
