# I've tested 11 video upscaling methods — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/426d25c89571502d

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-28 00:50 ingested 2026-10-29 02:40

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xha6jy/ive_tested_11_video_upscaling_methods/)

## retained text



```
I am not sharing any AI video, not promoting anything, these are all my genuine findings, and it's comparing of video upscaling models. That's why it is naming so many brands. I'm genuinely interested in others experiences with different video upscaling methods and hopeful that some of the bits of my post could be useful to someone.

A month ago I started making AI videos with Seedance 2.5 and pretty quickly wanted them in 4K (not much prior image/video AI experience before). So I went through the most relevant options I could find: local, local paid (Topaz, I have an RTX 5090) and cloud. 11 models in total, plus their variants and different settings where they were exposed. Overall this whole research took me 2 days and cost me like a hundred bucks. These are my findings.

One note before you read: I only judged these for my specific type of video, realistic interpretations of anime characters in realistic interpretations of their world. What looks bad to me might be exactly what you want (some people might say that Topaz's Starlight v2.6/Astra are great for their realistic generations/real videos but I found its bad for my use-case).

**LOCAL (FREE)**

My hopes weren’t high here, but I wanted to try anyway. For image upscaling, one of the popular tiled Qwen 2.1 workflows running locally actually beat every cloud and Topaz model I’d tried, so I was curious whether anything local could at least do a decent 1080p.

**Wan 2.2 upscale workflow (ComfyUI)**  
Not very impressive. It didn’t really look upscaled compared to the original. It mostly generated different skin moles and such, while hair and everything else looked the same. It’s also extremely heavy: an 8 second clip from 480p to 1080p took 43 minutes, and that’s without spilling into shared GPU memory.  
Discarded: way too slow for what it gives, and with my VRAM it could never get to 4K anyway.

**H3 Video Upsampler (ComfyUI)**  
Usually 2-3 minutes for 480p to 1080p. Bad smearing and temporal issues, on par with or worse than SeedVR2. I tried the latest version with and without the LoRA, and denoise levels all the way down to 0.  
Discarded: the smearing, and like Wan it wouldn’t get to 4K on my VRAM.

**SeedVR2**  
About 6 minutes for 480p to 1080p. Unimpressive result, same as SeedVR2 for images.  
Discarded: just bad. (sorry if I hurt someone's feelings about this hah. It always didn't look good, subjectively, on my kind of images and videos)

*Edit:* I'm downloading SeedVR2 7B fp16, since I'd only tested the 3B Int8 model that the default Comfy workflow comes with. I'll try more settings and update the post if it's much better. > *Update:* Got the 1080p → 4K output with the heavier weights, and it's honestly not bad. If I was forced to choose, I'd take it over Starlight 2.6 or the ByteDance upscaler. It doesn't mess up colors, shapes or blur, doesn't oversharpen, and adds nice detail. It just has pretty noticeable smear/artificial blur around moving objects, whether they move relative to the camera or on their own. Maybe that's down to split latent, which I can't turn off because I don't have 96 GB of VRAM :D It's the same kind of problem that knocked out Magnific Precision for me. > > Settings I used: euler, simple, denoise 1.0, lab color correction, split latent on with overlap 4.

**RTX Video Super Resolution (NVIDIA’s nodes in ComfyUI)**  
I’d liked the “RTX Video Enhancement” toggle in the NVIDIA app on YouTube videos, so I gave this a go. Super fast: 37 seconds for 1080p to 4K, model loading included. But pixel peeping the original against the output, they looked identical, as if no upscaling happened. It seems made for cleaning up compressed streams, and clean Seedance footage has nothing for it to fix.  
Discarded: it doesn’t add any detail.

**LOCAL PAID (TOPAZ)**

I skipped Starlight Sharp, since sharpness in the traditional sense isn’t what I’m after. Render times are for 480p to 2K; 1080p to 4K (5 second clips) was about the same.

**Starlight Mini**  
About 15 minutes. This is the conservative one. It still upscales and adds nice clarity, less than Flux, while keeping blurred backgrounds intact.

The catch: locally it adds a horizontal line across the whole frame, about 1/10 of the height from the bottom. Below the line the picture looks blurrier, as if that strip wasn’t upscaled at all. It happens on every render, whatever the export format. The same clip through Mini on fal has no line, so it’s the app, not the model. It’s a known bug on the Topaz forum since November 2025, and 1.7.1 still has it for me. And no, it’s not VRAM; nothing spills into shared memory on the 5090. The only workaround I can think of is rendering a second, flipped copy and compositing the bottom strip from it, which is way too annoying.  
This one made my final two (more below).

**Starlight Precise 2.6**  
About 11.5 minutes. Hair and accessories look as good as with Flux, but skin gets kind of overblown, borderline not good looking. On a different scene (1080p to 4K) it broke blurred backgrounds in a really ugly way, even at the minimum sharpness of 1, and higher sharpness was worse. It has “Precise” in the name, but it pushes really hard.  
Discarded: on footage that’s already clean, it just oversharpens.

**CLOUD**

**ByteDance video upscale**  
I ran it five times with different settings: general and AIGC presets, standard and pro enhancement, high and medium fidelity. Surprisingly, given how good Seedance 2.5 is, and it's the same company, and I was using Seedance 2.5 inputs, for their video upscale, it was bad every time. The first try added black outlines around the body, head and hair, and was otherwise about as soft as Mini. Later tries had color fringing, oversharpening, shifted colors, and it even invented new details on the outfits. Medium fidelity was the least bad.  
Discarded: it changes too much, and no setting fixed it.

**Flux 3 video upscale (Replicate and Higgsfield)**  
The best one. Cleanest skin texture, not overblown or oversharpened, but it still adds the clarity I wanted. Hair and accessories are as sharp as with Starlight 2.6, and it doesn’t break blurred backgrounds. Side by side with my original, the colors look practically identical too. I ended up preferring Precise over Creative, because Creative seemed to add slightly more realistic texture to materials.  
Made my final two.

**Topaz Astra 2**  
After the first test I thought it didn’t invent as much overprocessed skin as Starlight 2.6 and kept blurred backgrounds better. Then I tried another scene and realized I’d oversold it. It has the same odd sharpening as 2.6; it breaks the blur a bit less, but still oversharpens it way too much. The SDR to HDR option also washed out the colors completely, nowhere near the original mood, and it costs a lot more.  
Discarded: same sharpening problem as 2.6.

**Starlight HQ (fal)**  
A middle ground between Mini and 2.6. It adds a bit more nice clarity than Mini, but you pay for it with the same ugly sharpening as Astra 2 and 2.6.  
Discarded: the sharpening.

**Magnific video upscaler**  
I started with Creative mode: creativity 0, then 50, then the Animation & 3D, Artistic and Realistic presets. Everything came out oversharpened, oversaturated and anime-looking, and the presets looked basically the same to me. In hindsight that’s kind of expected from a “creative” model. Some of it looked a lot like Starlight 2.6, and I can’t help wondering if it’s built on Topaz’s weights, but that’s pure speculation on my part.

Precision mode surprised me. In the previews at strength 60, it looked good: nice clarity and detail, more than Mini, without oversharpening textures or breaking blurred backgrounds. Sharpen at 100 looked bad, and smart grain at 100 added extreme film grain, as you’d expect. A full 5 second render at 80 was already too much and started oversharpening the blurred background, so I’d guess 50-60 is the sweet spot. But the full clip also showed the real problem: visible smear on fast-moving objects, and the glossy floor and reflections didn’t look as good as with Mini or Flux.  
Discarded: the smear on motion, which the short previews don’t show.

**CONCLUSION**

After all that, it came down to two.

**Flux 3 (Precise)** is the best looking by far: clean skin, real clarity, and blurred backgrounds stay blurred. The downside is that it’s cloud only and you pay per clip. A 5 second 1080p to 4K clip cost me about $3 on Fal, or 33 credits on Higgsfield (about $1.55–2.06 if you buy credits as a regular top-up, cheaper on the annual plans).

**Starlight Mini** is conservative. It adds a little less clarity than Flux, but it never breaks anything. Locally it’s unlimited with the Topaz Video subscription ($59/month), but that’s where the line bug is. On fal it’s clean, at about $1.30 per 5 second clip.

Right now I’m leaning towards Flux video upscale the most, but I wish I could use the Topaz's mini locally without the bug too...

I was surprised how bad most of the options were for this kind of footage. Video upscaling still feels like it’s in its infancy.

I’m very open to suggestions and other people’s experiences. Is there an really good video upscaler I missed? Has anyone gotten rid of the Mini line locally? And if you’ve found good settings for any of the ones I discarded, I’d like to hear it.
```



document fetched 2026-10-29 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1xha6jy/ive_tested_11_video_upscaling_methods
