“So, what exactly are the technical differences?”
EDIT: as excpected the pictures were compressed during upload, here is a link to download the files higher quality
https://www.dropbox.com/t/0mAR7RRY3JtmVJ5Y
First time posting here, hope the screenshots don’t get too compressed. I wanted to share my test results for upscaling an 8-second 1080p clip to 4K using different modules.
The following is the original snippet from both wide and close shot
I never posted on the forum before I hope you can download them to see more in details.
Specs: RTX 5090 / AMD 9950X / 64 GB RAM
Heads up for my comparison, I’ve chosen the scene from this show for the following reasons
A) the room is cluttered with items and a lot of colorful objects
B) detailed words and figures on shirt and objects.
C) Testing of face enhancement.
I’ve compared between Starlight Mini, Sharp, Rhe and Proteus for reference, I can’t do the Starlight original module as I don’t have credits and if I am not mistaken it’s cloud only.
lets start with
Starlight mini
Close shot
- it is more intensive and truth be told is amazing with figuring out details in the background, if you look in the back left side the building intercom panel (I was also confused why it’s in living room filming set but anyways) is way more visible than all the other modules, none of them were able to enhance it.
In addition looking at the face it did a great job enhancing/sharpening without loss of details - The text on the shirt is clearer and sharpened, you can clearly read it, the Nano tubes are detailed and you can se the enhancement catching the less visible lines in the back and catching it, and finally the very tiny head at the bottom left of the shirt was captured although wasn’t enhanced properly to what this original shirt picture shows it still managed to grab as much details as possible.
9/10 for close shots
Wide shot
- Smoothens everything out, enhancing the quality especially on the big text such as the whiteboards texts, the text is much more clearer now and readable
- doesn’t try to enhance text on smaller objects such as the textbook on the bookshelf
- great with face enhancement and again doesn’t try to overdo it
10/10 for wide shots
Starlight Sharp
Close shot
- right from the start you can tell the face is “sharper” with more details in the hair however the eyes suffer a bit, loss of improvement for the background buzzer item,
- the text on the shirt is sharper however!, AI is heavily more used here and some changes are done to the letters, the N is more rounded and the S turned into a 5, but more importantly the nano tubes were over sharpened that they lose their 3Dness not to mention the brain figure at the bottom left became a mosh.
7.5/10
wide shot
the wide shot on the other hand is a mess.
- the whiteboards are blurry now and lines + text is woven together.
- Text on books completely wiped out and feels off
- face enhancement failed for further away, I see features but worse than other modules.
3/10
which bring me to the biggest difference between Mini and sharp… and in reality the other modules as shown under the difference in time. Sharp is much faster to process than Mini 6m47s vs 31m37s heck that’s 4.5X the time, but of course comes with downgrades.
Now are you going to start enhancing all your 1080p videos with Starlight mini? up to you but Personally I added comparison with Proteus and Rhea which are fraction of the time (same clip 24 and 40 secs) that’s almost 47.5X faster 47.5!!
THAT BEING SAID, in terms of actual improvements, Personally I’d say the starlight mini module is on par with Rhea for 1080p improvements, I never tried Starlight modules with 720 or lower videos so I can’t speak on that behalf.
Here are the Rhea and Prot modules for comparison.
Hope that answers your question 
8 Likes
I don’t think the Starlights were made for enhancing 1080p videos.
They were made for enhancing really bad lowrez videos. Like VHS and SD content and such.
1 Like
I was confused at first as well. Their announcement covers the differences. For starters “Starlight Mini” in the old “Topaz Video AI” is the same as “Starlight” in the new “Topaz Video”.
The model is also now a part of the Starlight “family” model selector, and loses the “Mini” designation, as it actually outperforms the base model in most scenarios. We will be building future versions of Starlight with new architectures over time - Starlight Sharp being the first in the series of next gen variants.
They described the new sharp model as such:
- New AI Model Variant: Starlight Sharp: One of our most powerful diffusion-based video restoration models gets even more detailed. Starlight Sharp can restore with an extra level of crispness not present in any other offering. Super demanding, currently available on Windows NVIDIA platforms for local rendering, or on the cloud.
It’s confusing because the sharp version looks to be FASTER than the mini.
Also: how is she sharp compared to the original (cloud only) SL
1 Like
Also: if i remember correctly - the original SL could process interlaced videos - sharp can’t?
Thank you tamer.aladdin, I appreciate all the work you put into presenting this. Very interesting!
1 Like
thank you 
Here is the docs page write up for both Starlight (formerly Starlight Mini) and Starlight sharp:
Hi, I would like to do some tests with Starlight and 35GB or more GPU Ram usage to see if there is a difference. Your source would be perfect footage for this, could you upload a few seconds of this source? Really only a few seconds to post afterwards here some screenshots of my results
Been seriously considering a new Nvidia GPU purchase, specifically for Starlight enhancement. A 5090 is definitely out of reach (the possibility of melting power cables are keeping me away from the high end anyway), but my options are between the 5070 Ti and 5080. Not really sure the extra £400 I need to spend on the latter is worth it though. Hoping the speed improvement over my existing 4070 Ti Super will justify the cost.
I would wait for the RTX 5080 Super if it is not urgent to upgrade. I wait out with my RTX 3070 as reportedly the new cards will come with 24GB of VRAM which would be very useful going forward considering Starlight Sharp already needs 16GB Vram as minimum.
Also reading through the Video AI topics for the devs it is a real challenge to shrink the most advanced models to 16GB Vram hence I would also wait out and get the higher VRam card.
1 Like
I do have an RX 7900 XTX in my other rig with 24GB of VRAM, and i’m really hoping Starlight Sharp eventually becomes available on this card. Annoyingly, I’m stuck to using QuickTime ProRes on this GPU as it’s not possible to choose 10bit H.265 as the preferred output. Another problem with the 5000-series Super cards are the inevitable battle with scalper bots on launch day!
I’m sure there is never a melting issue when you undervolt the 5090 @850mV as I did. My connector gets “hand warm”, power draw on the 12V connector keeps always below 400W. Ok speed droped around 8%, but TDP droped more than 30% and this makes the card highgly efficient.
2 Likes
I’ll probably need to upgrade my old Corsair 850w PSU in the event I do upgrade to a 5090, but I can probably get two cheap 5080s for the same price. Was going to wait until the Super cards drop next year, but obtaining one may be an exercise in futility as I’ll inevitably get beaten by scalper bots!
yes 850W psu is to low, but there is no need for a 1600W psu some did, specially not when you undervolt the gpu. I have the Corsair RMe 1200W ATX3.1, 2025 model
PSUs have the highest efficency around 50-60% usage, and that’s what i have while running Topaz, then draw is around 500-600W. seems enough for a 850W psu, but because there are short power peaks you can’t observe, a 850W psu is not enough.
2 Likes
Have been reading in another forum post about how good the RX 9070 XT is in Topaz Video. Have one in my livingroom PC, so I may swap the two cards over and give it a go. If it’s significantly better tban the 7900 XTX I may just keep my existing setup until the Super cards drop next year.
2 Likes
Hopefully we can get those cards if we look around at out local retailers. I always do it this way and I could have almost always pre-order a card at Alza for example…cannot wait for the reveal of the new cards and some performance improvement demos.
1 Like
Well, I finally got round to swapping my 7900 XTX for the 9070 XT. Biggest disappointment is the beta of Starlight now no longer works, and most of the enhancement models are slower than my 4070 Ti Super, with the exception of Gaia which seems to be significantly quicker, at least in the benchmarks. If I do use the 9070 for upscaling tasks, I’ll probably stick to purely animated or CG sources from now on.
well you say that I am rendering with V7 Beta, then I don’t close the program afterwards. This is starting to annoy me, I use the beta because the GPU takes a lot more Vram and Shared, I gives me bit smoother results. I can’t get rid of the impression that Starlight deliberately cuts the quality. Higher costs, in Studio no more 3x and 4x is bugy, more nurf than progression since Starlight release while their agenda says “Improving Starlight”, maybe it’s time to switch to SeedVR2…
2 Likes
