How do I upscale AI-generated video and images?
Match the upscaler to the subject rather than picking one and using it for everything. Fine structures — hair, fur, fabric weave — dissolve under a general-purpose upscaler and survive under one trained for them. Decide first whether you need a scale factor or a specific output resolution, because different models take different inputs. One well-chosen pass beats two generic ones, which compound artefacts.
Subject decides the model
Upscaling reconstructs detail rather than interpolating it, so what the model learned to reconstruct matters. Portraits, animal fur and detailed fabric need an upscaler tuned for fine structure; Aura SR is built for exactly that and runs at a fixed 4x. General images with broad shapes do fine on a general upscaler like Clarity, which is configurable between 2x and 4x. Using the general one on a portrait tends to smooth skin texture into plastic.
Scale factor versus target resolution
If you have a delivery spec — 1080p, 2160p — reason in target resolution and use a model that accepts one, such as SeedVR. If you simply want more detail than you have, a multiplier is easier to think in. Mixing the two mental models is where people end up upscaling twice, which compounds artefacts rather than improving the image.
Render high or upscale after?
Where a model outputs high resolution natively, that is usually better than rendering low and upscaling — native detail is real, reconstructed detail is inferred. MiniMax H3 outputs 2K natively, and VEO 3.1 and LTX 2.3 reach 4K. Upscale when the source already exists, when the model you need caps below your delivery spec, or when a cheap render plus an upscale costs less than the native high-resolution render.
Related questions
Build it on Floniks
Image, video, digital humans, and reusable workflows on one canvas. No card required.
Explore Floniks