Generative video in EXR: how to assess it for compositing

What has gone into the changelogs

Runway's API changelog holds four consecutive entries that matter to a comp room. On 20 August, HDR outputs in hdr10, hlg and sdr_rec709_10bit for Gen-4.5 and Aleph 2.0. On 31 August, HDR outputs can be delivered as scene-referred ACEScg OpenEXR sequences, in ACES 1.3 and 2.0. On 11 September, alpha channel in EXR. On 17 September, frame rate conversion.

Luma announced Ray3.2 in June with HDR and 16-bit EXR export, up to 20 seconds at 1080p and up to 16 keyframes within a single clip.

Lightricks, which maintains LTX-2.5 with downloadable weights, published between 8 and 15 September a set of LoRAs named after post-production tasks: Clean-Plate, Day-To-Night, Water-Simulation, Deblur, Decompression, Colorization, Ingredients, Slow-Motion and a spatial upscaler.

The change is not in image quality but in what leaves the service: layers instead of clips.

Why EXR can improve the workflow

OpenEXR can store high dynamic range data, multiple channels and precision suitable for compositing. ACEScg, when correctly declared and transformed, helps integrate elements into a managed colour pipeline.

But EXR is a container, not a quality guarantee. A sequence may originate from compressed video, contain a constant alpha channel or hold values that do not match the stated colour space. In Runway's 11 September update, alpha is preserved when it already exists in the source of the HDR process; the note does not promise an automatically generated matte for every generated video.

MP4 can still be suitable for references, editing or some final deliveries. It simply leaves less room for demanding isolation, highlight recovery and colour work than a suitable higher-precision source.

Read the licence before the quality

LTX-2.5 publishes a community licence allowing free commercial use for companies under 10 million dollars in annual revenue, with a paid agreement above that figure. It is a concrete, checkable condition, and it decides whether the model can enter a delivery.

Europe adds regional restrictions. The Gemini Omni documentation does not allow editing or extending videos uploaded from the European Economic Area, and Veo limits generating people in the European Union to adults and with permission. Several APIs acting as engines behind a service carry their own terms, worth reading one by one.

If the model you were using disappears, the problem is continuity rather than quality: how to prepare a replacement sets out the steps.

Check the delivery with a real shot

Use a shot from the real job, not the vendor example, and keep the discarded attempts too. The best clip in a gallery says nothing about how many tries your shot will need.

Record where every element came from

Do not assume that every output carries the same watermark or credential, or that the signal survives compositing. Inspect the file and retain its edit history, with a delivery note identifying the tool, version and date. The specific obligations are covered in how to disclose AI-generated content.

Context helps to calibrate the urgency. Netflix states in its second-quarter 2026 shareholder letter that generative AI workflows have been used in roughly 300 titles, with the largest concentration in post-production; Lionsgate took a stake in Runway in June; and SAG-AFTRA ratified that same month, with 91.4% of votes in favour, an agreement that further restricts the use of synthetics. The tool is already in the post room and the terms of use are being negotiated alongside it.

To decide whether a particular shot can be solved this way, tell us what source material you have and what has to be delivered.