Object Mask stops working until Premiere restarts as VRAM usage increases — reproduced on 8GB and 16GB GPUs in versions 26 and 27

Object selection repeatedly stops working during a masking session and only recovers after fully restarting Premiere. I have reproduced this with two different NVIDIA GPUs and in Premiere versions 26 and 27.

Comparison between GPUs

• RTX 3060 Ti, 8GB VRAM: typically completes around 1–2 clips’ worth of object masking before the warning appears.
• RTX 5080, 16GB VRAM: substantially faster and typically completes around 8–10 clips before the same failure.
• Premiere 27 lasted noticeably longer than version 26 in my testing, but eventually developed the same problem.

These counts are approximate observations, not fixed reproduction thresholds.

Steps to reproduce

1. Open my existing wedding-editing project.

2. Select a clip with multiple Lumetri Color effects.

3. Create Object Masks for selective colour corrections and track them through the clip.

4. Continue creating and tracking Object Masks on additional clips in the same session.

5. Monitor dedicated GPU memory in Windows Task Manager.

Actual result

Dedicated GPU memory rises sharply during tracking. Some memory is released afterwards, but the resting usage can also increase as the session continues.

Eventually Premiere displays “Object selection may be limited” and says the latest object-selection model is unsupported and a different model will be used.

Despite the fallback message, object selection no longer works for me. Dismissing the warning can release GPU memory, but does not restore the tool. Fully exiting and restarting Premiere restores it temporarily.

Memory observations on the RTX 5080

• Approximately 1.5GB total dedicated GPU memory usage with Premiere closed.
• Approximately 5GB attributed to Premiere shortly after reopening the project.
• Resting usage can rise to around 9–10GB during the session.
• Tracking peaks approach the card’s 16GB capacity.
• During some failures, total dedicated usage remained around 14.6–15.1GB while the warning was displayed.
• Usage sometimes drops substantially during a session, so it does not increase continuously.
• Closing Premiere releases the accumulated memory.

The repeated relationship between available VRAM, how much masking can be completed, and the warning suggests memory pressure or a GPU-resource-management issue. I would appreciate engineering confirmation rather than assuming this is a proven memory leak.

Expected result

Object selection should continue working across clips. If temporary memory pressure prevents an operation, the tool should recover when resources become available without requiring an application restart.

System and versions

• Windows 11; Reliability Monitor reports OS version 10.0.26200.2.
• AMD Ryzen 16-core processor.
• 128GB DDR5 system RAM.
• Current GPU: Gigabyte GeForce RTX 5080 Gaming OC, 16GB.
• Previous GPU: GeForce RTX 3060 Ti, 8GB.
• NVIDIA Studio drivers used during troubleshooting: 621.14 and 621.24.
• Premiere versions tested: 26.5.2 and 27. Exact version 27 build can be supplied.

Version 27 was tested using a separate copy of the existing project. I have not yet established whether a new, minimal project reproduces the same failure.

Additional system crash

During Premiere work on 12 November 2026, Windows also recorded LiveKernelEvent 141 at 22:04, followed by bugcheck 0x0000001E with exception 0xC0000005 at 22:06 after a black-screen incident. I have retained both dump files and can provide them privately. I do not know whether this crash shares the same cause as the masking failure.

Questions for Adobe

1. Is this a known Object Mask memory-management or model-loading issue?

2. Why does a model that works initially become unavailable until Premiere is restarted, even after GPU memory usage drops?

3. Is there a supported way to release Object Mask GPU resources without restarting Premiere?

4. Which logs or diagnostic capture would allow engineering to distinguish insufficient VRAM from retained resources or a driver interaction?

5. Is there a known fix or recommended build?

This is interrupting paid editing work. Please advise how to provide a project/sample and diagnostics privately for investigation.