For MiniMax H3 video, we hand it to the MiniMax H3 Latent Upscaler (Community). FlashVSR v1.1 still infers the sharpest new detail of anything we tested, and both are now on the Sogni Supernet — but FlashVSR shows temporal aliasing on camera dolly and pan moves, where fine detail crawls and shimmers between frames. We recommend it for scenes that are mostly static. If speed is your only priority, NVIDIA VSR was the quickest AI option in our RTX 4090 test.
We compared three open-weight approaches—FlashVSR, SeedVR2 and the MiniMax H3 Latent Upscaler (Community)—alongside NVIDIA's SDK and FFmpeg Lanczos. The task: upscale a 768p MiniMax H3 video while preserving its full picture and audio.
Here, 2K means 1440p: 2520×1440, preserving our source's aspect ratio.
All five options
| Upscaler | Why consider it? |
|---|---|
| FlashVSR v1.1 | On the Sogni Supernet. Best for mostly-static scenes. The sharpest inferred detail in our review, and it works from a finished video. On dolly and pan moves it aliases: fine detail crawls between frames. |
| SeedVR2 7B | Softer, more natural detail enhancement, but about 3.5× slower than FlashVSR in our 4090 test. The extra processing time was not worth it to us. |
| MiniMax H3 Latent Upscaler (Community) | Our pick for H3 video, on the Sogni Supernet. A community model that enlarges H3's saved latents, then refines the result with H3. Slightly softer than FlashVSR, but it holds together through camera movement. Needs more than a finished MP4. |
| NVIDIA RTX VSR ULTRA | The fastest AI option in our 4090 test. Uses NVIDIA's SDK. |
| FFmpeg Lanczos | Fastest overall in that test. A conventional CPU resize, without AI-generated detail. |
Compare the videos
Pick any two videos to play together, including the original 768p. MiniMax H3 Latent Upscaler (Community) was tested on a separate take of the same scene, so it lines up frame-for-frame only with FlashVSR and the original. Use 100% detail for faces, fur and edges; watch playback for flicker and unstable textures, and pay particular attention to the moments where the camera is moving.
Pick any two videos, including the original 768p, to play side by side.
Loading comparison videos…
Pause to compare the same frame. In detail view, drag either panel to pan both. Audio plays once.
Open any original video
Four video upscalers · RTX 4090
MiniMax H3 Latent Upscaler (Community) vs. FlashVSR · RTX 5090
Sharpness is not the whole picture
Judged a frame at a time, FlashVSR v1.1 wins. It invents more convincing fine structure than anything else here — fabric weave, hair, lettering, the edges of small objects.
Judged in motion, it gives some of that back. On shots where the camera dollies or pans, FlashVSR's new detail is not stable frame to frame: high-frequency texture crawls, shimmers and buzzes along moving edges. It is classic temporal aliasing, and once you notice it in a moving shot you cannot stop noticing it. On locked-off or near-static shots it does not appear, and FlashVSR is excellent.
The MiniMax H3 Latent Upscaler (Community) does not have that problem. It infers less new detail per frame, but it refines with H3 itself, so what it adds stays consistent across the shot and survives camera movement.
So the recommendation splits by content, not by score: FlashVSR v1.1 for mostly-static scenes, and the MiniMax H3 Latent Upscaler (Community) for MiniMax H3 video with a moving camera — which, for H3, is most of it.
How long does each take?
These are complete processing times for a 6.6-second clip, including model loading from disk into GPU memory and video export. Source-video generation and model downloads are excluded.
On the RTX 4090, Lanczos finished in 29 seconds, NVIDIA VSR in 40 seconds, FlashVSR in 3 minutes 6 seconds, and SeedVR2 in 10 minutes 48 seconds.
On the RTX 5090, the MiniMax H3 Latent Upscaler (Community) with refinement took 7 minutes 54 seconds. FlashVSR processed the matching video in 1 minute 56 seconds—about 4.1× faster. H3 used less peak GPU memory: 20.2 GiB, versus FlashVSR's 26.8 GiB. Both have since been added to the Sogni Supernet.
What does a 5090 buy you?
In our separate FlashVSR hardware comparison, the tuned 5090 had a median time of 1 minute 45 seconds, versus 3 minutes 6 seconds on the 4090: about 43% less waiting.
The MiniMax H3 Latent Upscaler (Community) method
This community method applies a learned upscaler directly to saved H3 latents, then refines the result with H3. It uses open weights, without a third-party upscale API.
It needs saved latents and conditioning—the generator's intermediate data. Retain those files and you can run upscaling as a separate job. The other four methods can start from a finished video.
We tested H3's learned upscaler with refinement. The version that skips refinement remains untested.
It holds up even at Turbo
You do not need H3's slow tiers to get a good result out of this path. The two clips below are FastH3 Turbo — the four-step engine Sogni Web uses by default for Turbo — enlarged and refined by the same community upscaler, and we think they are superb. They are the same two takes featured on the MiniMax H3 model page, where this path now ships as two-stage 2K: a Video Size choice in FastH3 settings that delivers 2688×1536 for 16 Spark ($0.08) per output second, the same price as Standard 768p.
The verdict
Both are on the Sogni Supernet, and which one you want depends on your shot. FlashVSR v1.1 infers the sharpest detail, upscales any finished video, and was about 4.1× faster in our 5090 test — use it when the camera is mostly still, because on dolly and pan moves its detail aliases. For MiniMax H3 video specifically, we hand it to the MiniMax H3 Latent Upscaler (Community): it works from H3's saved generation data, stays stable through camera movement, and looks superb even on a FastH3 Turbo render. These results cover short scenes at 1440p and one full-length clip at 2688×1536; they do not establish the best choice for every video or for 4K.
