Upscaling Generated Image-to-Video Footage: Tools, Fixes and Artefacts

You run a 1280×768 clip from Runway Gen-3 Alpha (2024) through Topaz Video AI hoping for usable 1080p footage. The output looks sharper. But the subject's face now pulses every third frame, the background slides one pixel left then jerks back, and a texture barely visible on the original has bloomed into a persistent checkerboard across the sky. This is not a failure of the upscaling model. It is the normal outcome when you upscale AI-produced video without understanding the artefacts already baked into the latent-space compression. The fix starts with knowing which artefacts the generation model planted and which upscalers amplify them.

Every image-to-video model produces a native output resolution lower than the file's pixel dimensions. A model that writes a 1920×1080 file may have an effective resolution closer to 1280×720 because the latent space compresses spatial information before decoding to pixels. This is the Claimed_Vs_Real gap: one full resolution tier. The effective detail in a Kling 1.6 (2025) native 1920×1080 clip is roughly 720p per frame. The effective detail in a Luma Dream Machine (2025) native 1360×752 clip sits lower still. An upscaler that quadruples the pixel count is multiplying hallucinations, not detail.

The practical effect: a frame-by-frame upscaler such as FFmpeg with Lanczos resampling preserves sharp edges but adds no information. It gives you a bigger version of the same distorted frame. A neural upscaler such as Topaz Video AI or Real-ESRGAN tries to invent the missing pixels, and that is where the trouble starts. A generated frame already contains latent-space artefacts: the model's encoding noise, the temporal attention leakage, the decoder's interpolation errors. These look like subtle grain or compression artefact at native resolution. An upscaling model treats them as real detail and enhances them.

Topaz Video AI (version 4.x, 2025) offers multiple upscaling models: Artemis for sharpening, Gaia for face recovery, and Proteus for general enhancement. On filmed footage these models produce clean results. On synthetic footage they introduce hallucinated detail because the source image's texture is already artificial. A generated cloud bank at native resolution may contain a subtle repeating pattern from the latent-space decoder. Topaz's Artemis model interprets that pattern as a real texture and amplifies it into visible banding across the sky. Gaia recovers facial features by matching them to training data, not to the original subject. A face that drifted during generation gets replaced with a sharper face that belongs to someone else.

Flicker Amplification

Generated video flickers because the latent-space decoder does not decode each frame consistently. Frame 47 may decode with slightly brighter shadows than frame 48. At native 24 fps this flicker is often invisible. After Topaz sharpens each frame independently, the difference between frame 47 and frame 48 becomes obvious: a pulsing brightness the human eye tracks immediately. This is flicker amplification, and it is the single most common complaint from users who upscale generated footage with Topaz Video AI. Apply temporal smoothing before or during the upscaling step. Topaz Video AI does not offer a temporal-aware upscaling mode as of the 4.x release (2025).

Real-ESRGAN (2021, maintained periodically through 2024) is the standard neural upscaler for open-source pipelines. Applied frame-by-frame to a generated clip, it produces the worst flicker of any common upscaler. The model is trained on photographic degradation, blur, noise, compression. Not on the smooth but inconsistent textures of a latent-space video. A single frame of Stable Video Diffusion (2023) output at 1024×576 contains a fine noise structure from the diffusion process. Real-ESRGAN reads this as grain detail and sharpens it. The next frame has a slightly different noise structure, and Real-ESRGAN sharpens that differently. The result is a 3-5 Hz flicker on every static surface in the frame.

The workaround with Real-ESRGAN: apply it only to keyframes and interpolate the rest, or run it through a temporal-consistency filter such as FFmpeg's tblend or a custom Python script that averages the difference between frames before and after upscaling. These workarounds add 2-4 GB of peak VRAM usage because the pipeline must hold multiple decoded frames in memory simultaneously.

FFmpeg (latest stable release, 2025) is the universal post-processing tool for generated video. It is the one tool that does not introduce its own neural artefacts. The standard command for frame-by-frame upscaling uses the Lanczos resampling filter:

ffmpeg -i input.mp4 -vf "scale=1920:1080:flags=lanczos" -c:v libx264 -crf 18 output.mp4

This preserves the source's sharpness without inventing detail. For generated video this is often better than neural upscaling because it does not amplify latent-space artefacts. The trade-off: the output looks soft because the effective resolution of the source was lower than the target. That softness is honest. A neural upscaler would replace it with fake grain.

For temporal consistency, add a tblend filter before the scale operation:

ffmpeg -i input.mp4 -vf "tblend=temporal,scale=1920:1080:flags=lanczos" -c:v libx264 -crf 18 output.mp4

This averages each frame with the one before it, reducing flicker at the cost of slight motion blur. On a clip with fast camera movement the blur is visible. On a static or slow-panning shot it is a net improvement.

Frame-by-frame upscalers treat each frame independently and worsen temporal inconsistency. Temporal-aware upscalers consider motion vectors across frames and preserve the original's temporal coherence. As of 2026, no commercial tool offers a dedicated temporal-aware upscaling mode for generated footage. The closest option: use a video interpolation model such as RIFE or DAIN between the upscaling pass. These force each upscaled frame to match its neighbours' motion.

Running Temporal-Aware Upscaling on Generated Clips

The pipeline runs: native generation → frame extraction → frame upscaling with Real-ESRGAN → frame interpolation to 2× or 4× → re-encode at original frame rate. The interpolation step smooths the flicker by generating intermediate frames that average the upscaling artefacts. This adds significant VRAM requirement. A 14B Wan 2.1 (2025) pipeline plus RIFE interpolation needs 16-20 GB of GPU memory. The result is usable for slow camera moves but introduces motion artefacts on fast action. For most generated clips, which already suffer from limited motion coherence, this is acceptable.

A simpler alternative: produce output at the highest native resolution available. Kling 1.6 (2025) outputs 1920×1080 natively. Veo 2 (2025) outputs up to 4096×2160 natively. If you can generate at 1080p, you may not need to upscale at all. If you must upscale, stay within a single resolution tier: from 720p to 1080p, never from a low base to 2160p.

Upscaler	Flicker Introduction	Hallucinated Detail	Temporal Consistency	VRAM at 1080p	Best Use
FFmpeg Lanczos	None	None	Preserved	0.5 GB	Clean upscale, no artefacts added
Topaz Video AI Artemis	High	High on textures	Degraded	6 GB	Only on low-motion clips with manual temporal smoothing
Real-ESRGAN frame-by-frame	Very High	High on noise	Degraded	4 GB	Only with keyframe interpolation pipeline
RIFE interpolation after Real-ESRGAN	Low	Moderate	Improved	12 GB	Best for slow pans and static shots
No upscale (native output)	None	None	Original	0 GB	Best option if native resolution meets requirements

AI upscalers introduce three classes of artefact to generated video that they do not introduce to filmed footage. First, flicker amplification: the frame-to-frame brightness and texture variation inherent in latent-space decoding becomes visible after sharpening. Second, hallucinated detail: the upscaler invents texture where the generation model produced noise, creating checkerboard patterns, false grain, and face swaps. Third, motion vector confusion: a temporal-aware upscaler misreads the generation model's soft motion boundaries as real edges and introduces ringing artefacts around moving subjects.

These are not bugs in the upscaling tools. They are the predictable consequence of passing a latent-space signal through a perceptual upscaler trained on photographic degradation. The generation model's output is not a photograph. It is a statistical reconstruction from compressed noise, and every upscaling pass that tries to add detail will amplify the noise as if it were real.

This suits the solo filmmaker who needs to squeeze a usable 1080p clip from a 720p generation, the open-source tinkerer running Stable Video Diffusion (2023) or Wan 2.1 (2025) on a consumer GPU, and the content creator who wants to know why their upscaled AI clip was rejected by a platform's quality check. It also suits the compliance person who needs to understand what is and is not fixable in generated footage before clearing it for broadcast.

Skip this if you are generating text-to-video only, if you are shopping for an all-in-one video editor with built-in generative features, or if you need a current price comparison for upscaling services. Prices change monthly and this site does not track them. Skip it if you are generating talking-head avatars. Those have different artefacts and different upscaling requirements.

The thing that most often goes wrong: a filmmaker runs a generated clip through Topaz Video AI expecting a Hollywood-grade clean-up, then blames the generation model when the result flickers. The generation model's artefacts are the starting condition. The upscaler's artefacts are your own addition. Know which is which before you click render.