Developers can apply DLSS 5 separately to characters and individual objects
When NVIDIA announced DLSS 5 a few months ago, the company faced strong criticism over images that appeared heavily AI-generated. NVIDIA responded by saying that DLSS 5 would preserve artistic intent and give developers more control over the final image. At the time, however, it did not explain how those controls would work.
At SIGGRAPH, NVIDIA demonstrated the tools for the first time. The company confirmed that DLSS 5 currently includes three models , while developers can select different models for individual scenes. They can also adjust the effect on separate characters, objects and other elements visible on screen.
The company demonstrated multiple models, global intensity settings, automatic character masks and engine-defined controls for individual objects. DLSS 5 adds a generative stage after the game produces a conventionally rendered frame. The original renderer still defines the geometry, camera, lighting setup and composition. DLSS 5 then processes that frame to improve materials, shadows, reflections and other visual details.
© NVIDIA 
© NVIDIA NVIDIA outlines three DLSS 5 challenges
The first challenge was preserving the developer’s artistic intent. Generative models can modify details that should remain unchanged, such as facial features, scars or costume elements. NVIDIA uses the rendered frame and internal engine buffers, including albedo, surface normals and lighting data, to keep the generated result connected to the original scene.
The second challenge was temporal stability. Video generation models usually process multiple frames together, but a game must respond immediately to player input. DLSS 5 operates causally, processing one frame at a time without looking at future frames. NVIDIA uses motion vectors and other engine data to reduce shimmering, drifting and swimming between frames.
The third challenge was performance. NVIDIA distilled knowledge from larger generative models into a smaller one-step pixel-space diffusion transformer. The model is designed only to improve the appearance of rendered game frames, rather than generate unrestricted images or videos. It appears that NVIDIA is still confirming that DLSS 5 will run on a single GPU with ‘efficient’ VRAM allocation. They did not go into details, so it’s not clear whether all 3 models will run across entire RTX 50 GPUs stack.
Developers can select different DLSS 5 models
DLSS 5 will include multiple models trained with different parameters. NVIDIA demonstrated three examples identified as Model A, Model B and Model C, with each producing a different interpretation of the same rendered scene.
© NVIDIA 
© NVIDIA A developer will not need to use one model across an entire game. NVIDIA says developers can select different models for individual scenes, environments or cutscenes. Models can also be mixed throughout the game based on the intended visual result.
Developers will have separate structure intensity and tone intensity controls. Structure intensity changes higher-frequency details such as ambient occlusion, reflections and subsurface scattering. Tone intensity applies broader changes to lighting, color and the overall appearance of the scene.
DLSS 5 can be controlled per object
DLSS 5 also supports automatic masking through the model itself. NVIDIA demonstrated the model identifying a character and applying the effect only to that character while leaving the surrounding environment unchanged. The effect strength can then be adjusted independently. I think this is pretty cool, given user feedback, some developers may, for instance, never apply any DLSS5 to characters.
Developers can create additional masks inside the game engine for individual objects or groups of objects. NVIDIA demonstrated separate controls for bottles, grapes, a pitcher and other props. Each mask can have independent structure and tone settings, allowing developers to disable DLSS 5 for some objects while applying stronger processing to others.
NVIDIA says DLSS 5 is intended to extend existing rasterized, ray-traced and path-traced pipelines rather than replace them. The company describes generation as a third neural rendering category alongside reconstruction and function approximation. DLSS 5 is scheduled to ship this fall, while NVIDIA continues to adjust its controls based on feedback from developers and artists.
Note: I used Gigapixel to increase the resolution on the slides from the video.
Source:
