Yet not tested properly, major bug in subtitles

Alex_Videoman

AS usual, nothing has been tested here and you rely on users to do your QA
If the movie has more than one audio tracks you can not select anything else but the first one, just UNACCEPTABLE

sismotik21

Hey friend, how are you? It’s been over a year and Unifab’s UPSCALER is still awful. With plant scenes and camera blur, the results are terrible; it looks ugly and unprofessional, with completely unrealistic faces. The only thing I’ve found that gives acceptable results is Vellum’s UPSCALER. The results are acceptable, but not professional; it’s overly exaggerated, like the characters are made up. It’s acceptable to look at, though, and it’s the only good thing I’ve found. Even though they’ve supposedly fixed it, nothing has changed; everything is still the same.

sismotik21

I just tried another video and UPSCALER EQUINOX. The result is really bad; everything looks unreal, very plasticky. It’s not a professional result, and it adds things that aren’t in the original film. Now I tried UPSCALER VELLUN, and this one gives a better result—a cleaner result, more identical to the original. But the images still look unreal. I’d give VELLUN a 7 out of 10. I hope UNIFAB improves; the flaws have persisted for years. I hope they take this feedback as a step toward improvement and strive for a professional result where the image is more realistic and not over-processed.

Ctrl+Z

@sismotik21 Thanks for the detailed feedback. To help our R&D team pinpoint the issue, could you share a few more details:

UniFab version number — it’s shown directly in the top-left corner of the client window. Please paste the full version string.
Which model exactly gave the bad result? From your description it sounds like the “plasticky look / unreal faces / content not in the original” mainly happens with EQUINOX — is that correct? Just want to confirm so we don’t chase the wrong model. (And VELLUN is the one you rated 7/10 as acceptable, right?)
Source resolution and output resolution (e.g. 1080p → 4K, 720p → 1080p, etc.).
If possible, could you share a couple of comparison screenshots (original vs. enhanced), ideally circling where the “plasticky / unreal / added content” artifacts show up? One or two frames is enough.

With this info we can reproduce the problem and escalate it to the model team. Thanks again!

sismotik21

Ctrl+Z Exactly, EQUINOX gives very unrealistic results, VELLUN gives results more faithful to the original, but there are still some issues with the characters’ faces, like excessive makeup, making them look like animated characters in some scenes. That’s why I gave it a 7 out of 10.

Ctrl+Z

@Alex_Videoman Thanks for the feedback and for catching this. We’ve confirmed the bug — in videos with multiple audio / subtitle tracks, the AI Subtitle feature indeed can’t switch to anything other than the first track. The issue has been handed over to our engineering team and will be fixed in the next release. Really appreciate you bringing this to our attention!

Alex_Videoman

Ctrl+Z this is basic testing again, this just outlines what I have been saying for 18months, you are not testing releases at all before and basically, we have to do it for you, UNACCEPTABLE and even more when you censure me for saying it
This is at least the 20th basic bug I report and you have the audacity to treat me like this shame on you

sismotik21

Alex_Videoman I opened a thread, I hope it’s not censored, but there are things that UNIFAB does well and things that are a disaster, ruining the result. I really hope the people at UNIFAB come across the thread I opened and can finally correct this error. I took the time and effort to point out their mistake; they should appreciate that, it’s for their own improvement. Who else is going to take the time to point out their mistakes for free so they can improve? They should reward those users who find errors and take the time to open a thread. UNIFAB, you should reward the users who help you with your errors.

Ctrl+Z

@Alex_Videoman The multi-audio track selection issue has been fixed in the latest version — you can now switch to any track, not just the first one. Please update and give it another try.

On the broader point about testing: we hear you, and we understand the frustration behind 18 months of flagging issues like this. The fact that this kind of basic functionality gap made it to release is a fair criticism, and it’s something we need to address internally — not just acknowledge. We’re taking that back to the team.

sismotik21

Ctrl+Z I will create a thread to upload the results, and the only real reason is to improve UNIFAB, to make it a product on par with the best professional AI.

Ctrl+Z

sismotik21

On your concern about the new thread: our goal is to keep this community a constructive and respectful space. Genuine feedback that follows the community guidelines is always welcome, and we’d ask everyone to help uphold those guidelines together so the valuable content gets the visibility it deserves.

Thank you for taking the time to put together the dedicated comparison thread — we’ve seen it, and it’s a really thoughtful, useful piece of work. I’ll head over to that thread and respond case-by-case on the specific samples you posted.

We also fully understand how frustrating it is to keep flagging the same class of issue for a long time without seeing it fully resolved. Model iteration objectively has a longer cycle than regular bug fixes, but your material will be brought into the next round of training-sample evaluation. Once the model team has a conclusion, I’ll come back to this thread and update you directly.

One last thing: we genuinely value constructive feedback like yours, and would ask for a bit more patience as the models mature. For users who consistently contribute high-quality feedback, we run occasional official recognition / rewards — we’ll reach out when that applies.