Oscar winner Ben Affleck says in a closed-door meeting that public video models address physical literacy, while commercial barriers are thawing open-source weights
Oscar winner Ben Affleck said in a closed-door meeting that public video models address physical literacy, while commercial barriers are thawing open-source weights, and then using the crew's own live-action master footage to train the final layer of the film. Netflix has cash-acquired his post-production AI company InterPositive, founded in 2022, for approximately $587 million, as disclosed in SEC filings.
The acquisition brings in 16 engineers, researchers, and creative personnel, with Affleck serving as a senior advisor. He stated that public model footage cannot enter the film industry: camera movements lack breath, lighting lacks dramatic tension, and multi-camera setups are incoherent. Moreover, models trained on peer or unauthorized materials cannot pass Hollywood legal reviews.
The approach he described does not involve pre-training from scratch. Open-source weights have already learned color, gravity, and basic three-dimensional motion from a large amount of video. InterPositive thaws the weights and continues training from the last stopped step, focusing computational power on the final film layer. For this layer, the company self-funds and has filmed approximately 8 months of private data in a controlled studio, aligning with professional cameras, lighting, color temperature, and grain, without using web-scraped videos.
The workflow involves shooting and training simultaneously. Each new film fine-tunes a dedicated model using footage processed on the same day, with confidential materials not uploaded to third-party large model companies. Data remains with the crew, and iterative results belong to the creators currently filming the project. Upon completion, they receive the finished film and a private model that has learned the visual style of the film. The tools focus on missing shots, changing backgrounds, incorrect lighting, wire removal, re-lighting, and reconstructing images, rather than text prompts.
Netflix co-CEO Ted Sarandos stated in the Q2 earnings call that generative AI has begun to impact hundreds of productions, with about 300 works using some form of generative tool this year. The model from InterPositive is required to work with Netflix's in-house visual effects department Eyeline and third-party tools. The $587 million divided by 16 people amounts to approximately $36.7 million per person in cash consideration, which is the arithmetic of the acquisition, not a valuation per person.
This is an acquisition of content assets, not a subscription to model interfaces. The buyer is Netflix, and the sellers are Affleck and the investors of InterPositive, with cash consideration of approximately $587 million. The money buys private stage data, post-production processes, and a team of 16 people, not the usage volume of public video models. The beneficiaries are the crews that keep daily sample footage on set and the streaming services that need to reduce visual effects costs; the ones under pressure are those relying on web-sourced materials and charging for single generative outputs.
Source: Public information
ABAB AI Insight
Affleck quietly founded InterPositive in 2022, publicly stating that AI cannot write Shakespeare or be an author, only acknowledging its craftsmanship. The company shoots proprietary data on a closed soundstage, with the first version of the model designed to handle visual logic, editing continuity, missing shots, changing backgrounds, and incorrect lighting. In 2024, he mentioned in a CNBC conference that cost reductions would allow more people to afford filmmaking. In March 2026, Netflix acquired the company, with the cash consideration noted as approximately $587 million in the July 10-Q.
The money stays within the production chain. Daily sample footage does not go to third-party large model companies, and fine-tuning results follow the film. After Netflix's acquisition, the team is integrated into its production system and is required to ensure compatibility with Eyeline and external post-production tools. Sarandos includes generative tools in the cost items of hundreds of productions, with about 300 works using them this year. Public models continue to sell single generative outputs, while InterPositive sells the private weights left after a film is completed.
This is similar to Pixar keeping the rendering pipeline on set and game companies fine-tuning action models using their own motion capture. Hollywood is in a control phase: public video models can already produce images, and the added value lies not in retraining a general base but in who controls unpublished master footage and the training boundaries that can pass legal reviews.
Structurally, this is a reconstruction of the industry chain. The general layer is handled by open-source video models, while the film layer is injected with the crew's master footage. Once pre-training data is mixed with peer films, the finished product cannot pass copyright review; private sample fine-tuning transforms the model from a callable interface into an asset that follows a specific film. Computational power can be purchased externally, but 8 months of controlled stage and daily processed footage cannot be extracted from public models.
ABAB News · Cognitive Laws
General knowledge can be open-sourced, while style remains in the master footage.
Single stunning outputs do not connect seamlessly.
Not uploading sample footage ensures the model belongs to the crew.