AI upscaling and frame interpolation for archviz animation, without the plastic look

Every archviz studio that does animation eventually runs the same math. The client wants 4k at 30 frames per second, the render engine wants a week per shot, and the deadline is on Tuesday. AI upscaling and frame interpolation have made that math workable in a way that was not possible three years ago, but they come with tradeoffs that need to be understood before you promise a client a delivery date.

The first tool worth naming is Topaz Video AI, currently on version 6. It is a temporal AI upscaler that takes rendered animation sequences and upscales them to a higher resolution while removing residual noise and stabilising temporal artefacts. On archviz animation the results are genuinely useful. A 1080p render taken to 4k with the Rhea model is visibly cleaner and sharper than a native 4k render at the same quality settings, and takes about a fifth of the render time per frame.

The tradeoff is that Topaz occasionally hallucinates fine detail. On brick facades it can invent brick patterns that do not exist in the source render, and on foliage it can produce leaf shapes that were not in the original. On most archviz animation, the client will not notice, but on close up hero shots of a specific facade this behaviour can push you off brief. The mitigation is to render close up hero frames at native resolution and use Topaz only for medium and wide shots.

Frame interpolation is the other side of the same problem. If you render at 24 frames a second and the client requests 60 for a smoother feel, RIFE 4.9 or the newer Topaz Chronos model can interpolate the intermediate frames from your rendered sequence. This is qualitatively different from upscaling because you are inventing time, not detail. RIFE is the open source option and runs on any modern GPU. It is genuinely good for camera moves and static subject animation.

Where interpolation fails is when there is fast object motion in frame, for example a car driving across an intersection or a rendered person walking past camera. The interpolated frames can produce ghosting or double image artefacts around the moving element. For archviz that is usually manageable because the camera is the main moving element and the buildings are static, but if your shot has a busy intersection, you will need to mask out and hand render the problem regions.

DLSS 4 sits in a different category because it operates during rendering, not after. If you are rendering in Unreal Engine 5, Twinmotion, or Chaos Vantage with an RTX 40 or 50 series GPU, DLSS 4 lets you render internally at 1440p and upscale to 4k in real time. The quality is remarkable. For real time client presentation, DLSS 4 with the transformer model has effectively removed 4k as a performance question on capable hardware. For offline path traced output, DLSS is not part of the equation.

The workflow I run in the studio for a 60 second 4k 30fps animation looks like this. Render everything at 1440p 24fps in Unreal 5 with path tracing enabled. Upscale to 4k using Topaz Video AI Rhea. Interpolate to 30fps using Topaz Chronos or RIFE. Grade the final in DaVinci Resolve. Total render time for a shot that would have been six days at native 4k 30 comes down to about 30 hours. The quality difference is visible if you look for it but never enough to affect client acceptance.

There is a specific rule about AI processing that I hold to firmly. Do the AI work last, after the offline colour grade is locked. Running upscaling before grading means the upscaler is making decisions on ungraded footage that will shift under the grade, and you can get strange edge artefacts around graded highlights. Grade the source render at native res, then upscale, then do the final master pass in the timeline.

The hardware question is worth naming. Topaz Video AI benefits from a strong Nvidia GPU with plenty of VRAM. On an RTX 4080 with 16gb, a 4k upscale runs at about 6 to 8 frames per second depending on model. On an RTX 5090 with 32gb it runs at 15 to 20 fps. If you are doing a lot of upscaling professionally, the upgrade pays back in weeks. RIFE is lighter and runs well on any modern GPU including consumer cards.

Client communication around AI processed animation is worth being explicit about. Most clients do not care and never ask. Some technical clients ask specifically whether the output is native rendered or AI processed, and they deserve a straight answer. Naming the process in your production notes, for example rendered at 1440p and upscaled to 4k with Topaz Video AI, means you have said it before anyone asks. That is more defensible than presenting AI processed footage as native rendered and having a technically literate client spot the difference.

The last note is about future proofing your masters. Save the native resolution unprocessed render sequence alongside the AI processed final. In 18 months a better upscaler will exist, and being able to reprocess the source for a portfolio update or a re release is valuable. Storage is cheap. Delivering an animation and losing the source frames because the drive was reused is a mistake I have seen more than once, and it always costs more than the drive did.