Machine Learning Upscaling (What's New in 2026)
Indicates a feature suggested and voted up by users on the Flame Feedback portal.
On Rocky Linux, select the Machine Learning option in the Destination's Resolution Mode drop-down of a Resize & Crop, Render, Write File node or in the Media Export window to increase the resolution of a clip using machine learning upscaling models. The resolution of the clip can be doubled, tripled, or quadrupled by choosing the desired factor in the Scaling Presets drop-down.
Rendering Platform option
The Machine Learning inference can be run using either the CPU or GPU. Use the CPU / GPU drop-down button to select the mode of your choice.
Running the Machine Learning inference is faster on the GPU, but larger frames that cannot be processed on the GPU can be rendered using the CPU.
Caching a Machine Learning Upscale Engine (Rocky Linux only)
Contrary to the machine learning models available in the Timewarp and Morph tools, the ML Engine Cache function enabling the usage of the NVIDIA's TensorRT cache is enforced for the machine learning available in Resize & Crop. Because the cache requires a new tensorRT engine for each combination of GPU and image size, a new engine is generated each time the option is used on a new combination. Each model is stored in a cache in /opt/Autodesk/cache/tensorgraph/models and is used the next time the same GPU and image size combination is required, making it much faster to use the model on a combination for which you already have the cache.
Bit Depth option
An option to run a model in 16-bit fp or 32-bit fp is offered.
Selecting 16-bit fp offers two advantages:
- It allows running a model on higher resolution images, since running in 16-bit fp requires less memory.
- It provides better performance.
Improve the Playback & Scrub Performance for Machine Learning Upscale (Rocky Linux only)
Since a new inference of the Machine Learning upscaling model must be performed at each frame, the playback and scrubbing in Batch, Batch FX, and Modular Keyer are impacted.
Therefore, the application applies a simple Linear filter by default while playing back and scrubbing. Use the Viewport option of the Machine Learning Upscale preferences to modify the behaviour.