USC ETC finishes 8-bit generative AI short for theatrical film print

This development proves that the 'fragility' of generative AI—specifically the artifacts and limited bit depth of 8-bit web outputs—is no longer an absolute barrier to theatrical distribution. By bridging the gap between 8-bit SDR and 32-bit floating-point workflows, ETC is standardizing how studios can incorporate generative assets into high-end finishing pipelines without compromising existing exhibition standards. This shift positions AI-generated content as a viable primary source for cinema rather than just a reference or previs tool. Watch for whether major post houses adopt similar standardized 'restoration' headers for AI content arriving from non-professional generation platforms.

The USC Entertainment Technology Center's work on The Bends represents one of the first documented attempts to build a complete post-production pipeline specifically for generative AI content destined for theatrical exhibition. The project's technical approach inverts the traditional finishing workflow: rather than managing fidelity downward from camera originals, the ETC team systematically builds quality upward from 8-bit SDR sources. Sahil Lulla, Head of Innovation at Bitrate Bash and R&D lead on The Bends, described the 8-bit SDR source as functioning like a proxy that requires fidelity increases at each downstream step. The three-step Topaz Video AI sequence moves from Nyx denoising at source resolution, through Gaia upscaling to 4K 32-bit EXR, to Hyperion inverse tone mapping that expands the color gamut from Rec.709 to Rec.2020, covering 75.8% of the visible spectrum compared to the 35.9% of the original sRGB source.

The ETC has positioned The Bends as a proof of concept for broader industry adoption of generative AI in professional post-production. The project was presented at Infinity Fest, where the pipeline demonstrated that AI-generated footage could achieve 16-bit EXR output with extended grading latitude suitable for high-end colorist workflows. The ETC's approach addresses a specific gap: most generative video models output 720p to 1080p H.264 files that fall far short of cinema standards, and naive upscaling amplifies compression artifacts and banding rather than resolving them. The staged model selection process tested all available Topaz models including Proteus, Artemis, and Theia before settling on Gaia for its performance with organic textures and motion consistency.

Beyond the technical pipeline, the ETC's work on The Bends also tackled the creative control problem that has limited generative AI adoption in long-form storytelling. The project also experimented with projecting AI outputs onto 3D primitives, treating generative models as a texture engine for simple geometric shapes. These techniques address the stylistic consistency challenge that the ETC identified as the primary barrier to generative AI moving from experimental pitches and previs into full production workflows.