Google DeepMind developed custom versions of its Veo and Imagen models to support the production of the film project “Dear Upstairs Neighbors.” The technical team trained these specific builds using original concept art from the production to ensure visual consistency throughout the footage. This specialized approach attempts to move beyond the quality issues and visual “slop” currently produced by standard generative AI tools.
Key details
- The project utilized custom-trained builds of the Google Veo and Google Imagen models.
- Training data for these specific models consisted of original concept art created for the film.
- The film project is titled “Dear Upstairs Neighbors” and is associated with the Tribeca festival.
- The workflow prioritizes creator-led training data over the general-purpose prompts used in standard AI software.
Why it matters
This project signals that professional filmmaking may transition away from general-purpose AI models in favor of custom systems trained on a film’s specific visual identity. While current public models often produce inconsistent results that production houses cannot use for high-end features, fine-tuning on proprietary concept art allows for greater control over artistic style. This strategy addresses the technical limitations that have caused previous high-profile partnerships between studios and Silicon Valley firms to evaporate. Watch for more creators to demand private model training as a way to maintain visual standards while using automated tools.
Read the full story at The Verge
Leave a Reply