8/29/26

Vanishing services, six-figure bills and H.264 instead of EXR: why studios are moving generative models onto their own hardware

Interesting From The Web

The industry letter “Open-Weights and American AI Leadership”, signed by 35 participants including Nvidia, Microsoft, Meta, IBM and Hugging Face, has put the argument over open weights back on the agenda. Its authors talk about accessibility, competition and safety through open review of models by the community. On the ground of film and VFX production this argument looks different: there a model is judged by whether it will survive until the project is delivered, what its iterative use will cost, and in what format it hands over frames.

The publication fxguide examined the practical side of the question together with JD Vandenberg, a consultant on generative pipelines and color science who previously worked at Disney and Netflix. The conclusion comes down to this: open weights give studios version stability, control over the material, a cost with a ceiling on it and the ability to build generation into an already existing pipeline — at the price of technical expertise and in-house infrastructure.

Исчезающие сервисы, шестизначные счета и H.264 вместо EXR: почему студии переносят генеративные модели на своё железо — An image generated locally with the open Z-Image-turbo model (an open Diffusion Transformer). Image: Jack Higgins / fxguide

A model can disappear in the middle of production

In production, software is not updated without a reason — predictability matters more than fresh features. Generative models live by different rules: they are updated often, without warning and without a full list of changes. A frame, a video or a sound produced from the same input with the same seed can change if the provider does not offer versioned snapshots and you have not pinned a specific one.

The worst-case scenario is a service shutting down overnight. That is exactly what happened to the Critterz project after OpenAI wound down the Sora app and web service in March: the project missed its scheduled Cannes premiere, its co-producers had to look for a new AI partner, and the release was pushed to early 2027. The same announcement ended the story of Disney’s announced $1 billion investment in OpenAI and the related three-year character licensing agreement — the deal was public for three months and was never finalized.

Content security stands in the way of sending source material to someone else’s cloud

Production networks today are segmented, and outbound traffic is tightly controlled by MPA and TPN requirements. Sending source material to a hosted model and pulling the results back punches a channel straight through the very barriers content security exists for. Pre-release material travels down that channel, and legal agreements work as after-the-fact compensation rather than as control. At some point someone has to confirm that the provider’s infrastructure is capable of holding such material: either pay for an audit or trust a vendor that has never been through one. Studios have the mechanism — but only a handful of generative service providers have passed a TPN assessment. A local model works even on a fully isolated network, with no data handed to a third party.

The local option has a downside too, and fxguide does not hide it. ComfyUI, for all its merits, remains a source of risk: an attacker can slip a virus or a worm both into the program’s own code and through the open plugin ecosystem, and the volume of code is too large to be audited by an in-house IT department. Security specialists in VFX treat open generative builds as hostile by default — especially when artists use their own devices and tools they have downloaded themselves. The alternative is to assemble a trusted solution yourself, out of vetted open libraries.

The bill for iterations cannot be calculated in advance

Generative production is iterative, and it is still impossible to predict how many attempts a model will need to hit the brief. The short film Air Head, made in 2024 by the Toronto collective Shy Kids on Sora, is telling: post-production supervisor Patrick Cederberg estimated that roughly 90 seconds of the final running time grew out of “hundreds of generations of 10–20 seconds,” and put the ratio of raw material to edit at about 300:1. He himself noted that the count was approximate, and the figures do not add up exactly, but even on a conservative estimate it is more than 80:1. For live-action features 10:1–30:1 is considered the norm, and an experienced team knows how to estimate that ratio in advance.

As of August 2026, API rates hold in a range of roughly $0.05–0.75 per second of output. The cost of a finished minute is the price of a second multiplied by a poorly predictable coefficient. The software industry has already run into this: in June, a single employee at a San Francisco fintech company burned through $81,267 worth of tokens in a week while developing a game. In generative media the same cycle spins at a noticeably higher price per iteration. A spending limit only relocates the problem: the counter will stop in the middle of an episode on delivery day, and instead of finishing a shot the team will spend the night getting a budget increase approved. Local hardware costs what it costs, whether you did ten takes or ten thousand.

H.264 versus EXR: the formats do not match

The gap between what generative services hand over and what the post pipeline accepts comes down first of all to format. Most video models output H.264: 8-bit, 4:2:0, display-referred, most often sRGB. There is no alpha, no depth, no cryptomattes, no scene-linear data, but there is banding that survives any subsequent color grade. Cederberg noted that Sora did not allow rendering additional passes such as mattes or depth at all, and that Shy Kids generated everything at 480p and upscaled it with third-party tools.

A VFX delivery is, as a rule, EXR: a single 4K half-float frame across three channels weighs around 53 MB uncompressed, multilayer EXRs at 200 MB per frame are routine, which means 576 GB for a two-minute shot at 24 frames per second. At such volumes cloud storage and outbound traffic turn into a line item, and transfer time into a schedule constraint. A model in someone else’s cloud delivers the format that is cheapest for the provider to serve; a model on your own hardware can be placed inside your own color pipeline and made to write what the pipeline actually reads. The industry is solving the problem, but the gap is still wide.

Studios already have the hardware — it just needs to be pointed at a different task

Most video models ship in distilled or quantized variants that run on a single workstation. The trend favors studios: quantization and distillation cut VRAM requirements faster than growing parameter counts raise them. Vandenberg runs quantized and distilled versions of most current open video models on a single RTX 5090 — a consumer card with 32 GB of video memory. At the top end, the RTX PRO 6000 Blackwell with 96 GB retails for about $13,250; at launch it was noticeably cheaper.

The main thing is that studios, VFX houses and post facilities already own significant GPU capacity. Render farms and fleets of GPU workstations have already been bought — all that remains is to redirect them to a different workload. A colorist’s station with several cards running FilmLight Baselight or Blackmagic DaVinci Resolve sits idle outside working hours, even though it is suitable for inference.

From this follows the benefit that is hardest to demonstrate — integration. Your own model can be called from a Nuke node or a Resolve plugin, jobs can be sent to the same render manager with the same queue, priorities and logs as the rest of the show’s renders, color can be described once in an OCIO config or an ACES pipeline instead of being guessed at clip by clip, and versions can be tracked in the same asset management system. A hosted API can be scripted, but it stays outside the pipeline.

Open weights are not yet open source

Before building production on this, you have to sort out the licenses. Open weights mean the parameters can be downloaded, but they do not guarantee the right to fine-tune the model, distribute it or use the result in a commercial delivery. Open source in the usual sense is an OSI-type license permitting study, modification and distribution of the sources. Current open models for video and sound are scattered across that spectrum: some under Apache 2.0, some under community licenses with usage restrictions, at least one requiring a separate commercial agreement, and some allowed only for non-commercial use. A model that produces a great image under a license prohibiting commercial use cannot be put into a delivery, and conform is an expensive place for such a discovery.

Most of the leading open models in this area are made outside the US, and regulation of open weights is currently changing. Hence a practical piece of advice: archive the weights you actually use rather than pulling them from a hub at the moment a job launches.

Why this discussion has started right now

The argument about open weights in media production has probably ripened for a combination of reasons, and the letter is only its top layer. Closed models behind an API are not going anywhere: running things locally is non-trivial, and taking on your own infrastructure makes sense for far from every team. But several factors have coincided for studios: the GPU shortage and more expensive memory, because of which some mid-sized companies are heading to the cloud anyway; the accumulated experience of projects halted by a service shutting down; and the difference between how production is built and how a consumer AI service is built.

Against this background, growth in hybrid schemes familiar from the open software world is likely: the weights are freely available, while the money is charged for support, a commercial license and service. For small teams that would mean a more affordable entry point; for large facilities, a chance to bring generation back under their own quality and security control. The hardest hit would be providers of closed services that are prepared neither to pass industry audits nor to guarantee that versions stay unchanged.

Source

Save

5186 23

792

Votes: 5