Marianne doesn't paint what she sees—she paints what she remembers Héloïse being when no one was watching her.
The film's central architecture depends on this: the portrait emerges not from observation but from Marianne's interior archive of desire, glances, and half-understood moments reconstructed in solitude.
When Héloïse finally sees the finished work, she doesn't recognize herself—she recognizes Marianne's version of her, which has become the only version that matters because it's the one that will survive. Sciamma shoots memory sequences in soft focus, then sharper focus, then painterly tableaux, deliberately collapsing the boundary between what was observed and what was invented.
This is the alignment problem solved backwards. Generative AI struggles with alignment because it must choose whose desires to encode: the user's, the trainer's, the culture's. Portrait offers a model where that problem disappears entirely—there is no objective Héloïse to verify against. The painting is true because it satisfied the person making it. No one in the film questions whether Marianne got Héloïse 'right'—they only ask whether the work is beautiful, whether it moves them, whether it aligns with the emotional reality the artist intended.
Scale it. Industrialize it. Make it serve preference instead of person—and the arrangement collapses, for it cannot remain intimate and contained to one artist and one subject.
Rewatch the scene where Marianne shows Héloïse the finished portrait (around 1:23:00)—pay attention to Héloïse's silence and the single line she speaks. Then read the 2023 interview where Sciamma discusses why she chose not to show the audience the painting's surface in detail until that moment.
Find the video essay 'Painting and Narrative in Céline Sciamma' by Insider Film (2021), made before generative AI became mainstream, where the author diagrams how Sciamma structures visual information as memory rather than documentation—it predicts every formal choice the film makes and accidentally maps the architecture of diffusion models.