Another day, another saddness
still no beta for mac
5 Likes
Iâve identified two reproducible crashes in Topaz Video v1.5.0 that have prevented me from completing renders. Both occur with the bundled ffmpeg.exe. The same Aion 60fps workflow ran reliably for dozens of hours on the previous Topaz version and broke after the v1.5.0 update.
I have full diagnostic data, including 10 crash dumps; WinDbg !analyze -v output for representative dumps from each bug; a verbose ffmpeg log of a deliberate Bug 1 reproduction; and system info. Files uploaded to your target.
Bug 1 â Crashes within ~20 seconds of CLI launch using Aion (reproducible on demand)
Faulting module (visible): VCRUNTIME140.dll
Real cause (per call stack): videoai!AuthManager::ReloadInstance calls cv::copyMakeBorder in OpenCV with a bad pointer, OpenCV calls memcpy, which crashes on the null/invalid destination
Failure hash: {cafa3924-e068-da74-54ab-b234e71ce764} â identical across 9 separate reproductions
The license-check path appears to invoke OpenCVâs image-padding function during tile processing, which is unusual and looks like a code path that shouldnât exist
Bug 2 â Crashes after extended (within GUI) Aion renders (memory or resource leak)
Faulting module: onnxruntime.dll v1.23.2.0 (DirectML execution provider path)
Real cause: aiengine!aiutils::tzutils::verifyCompressedBuffer calling into ONNX Runtime DML, null pointer dereference deep in DML provider code
Failure hash: {c77c00a5-a265-9d9c-eba1-ea25491d117e}
Triggered after 17+ hours of GUI Aion rendering. Reproduced twice within 12 hours.
Why is the DirectML execution provider active during a CUDA inference workload on an NVIDIA GPU?
Environment
Topaz Video v1.5.0 (license active until 2026-06-25, user stevemit@twc.com)
Windows 11 Build 26100.2 (24H2)
NVIDIA RTX 5070 Ti, 16GB VRAM, [595.79]
96 GB RAM, 24 logical cores
Source: 1080p progressive H.264 CFR 29.97fps from clean Blu-ray rip
VC++ Redistributable 14.44.35211.0 (verified up-to-date)
Reproduction for Bug 1 (works every time, ~20 seconds to crash):
ffmpeg.exe -loglevel verbose -y -ss 00:00:00 -t 00:05:00 -i âsource.mp4â -filter_complex âtvai_fi=model=aion-1:slowmo=1:rdt=0.01:fps=60:device=0:vram=1:instances=1,scale=out_color_matrix=bt709,format=yuv420pâ -c:v hevc_nvenc -preset p4 -tune hq -rc constqp -qp 18 -pix_fmt yuv420p -color_primaries bt709 -color_trc bt709 -colorspace bt709 -color_range tv -fps_mode cfr -an âoutput.mkvâ
Bug 1 representative call stack:
VCRUNTIME140!memcpy_repmovs_Intel+0xb
opencv_world4120!cv::copyMakeBorder+0x99e
opencv_world4120!cv::copyMakeBorder+0x2dc
videoai!AuthManager::ReloadInstance+0x2a6df
videoai!AuthManager::ReloadInstance+0x36ff3
videoai!AuthManager::licenseType+0x162e
videoai!TiledProcessorRunnerPool::tileProcessorThread+0x17e
Bug 2 representative call stack:
onnxruntime!OrtSessionOptionsAppendExecutionProviderEx_DML+0xcc80
[12 frames inside onnxruntime.dll DML provider path]
aiengine!aiutils::tzutils::verifyCompressedBuffer+0x1925f
[aiengine internal frames]
msvcp140!Concurrency::details::_Task_scheduler_callback+0x49
Note on log warnings: Verbose ffmpeg log shows three âUnknown exceptionâ warnings on JSON file checks immediately before Bug 1 crash:
audio-codecs.json
model-recommendation-rules.json
video-encoders.json
These may be related precursors to the OpenCV/AuthManager crash that follows seconds later.
Iâm happy to provide additional repro data, run further diagnostic captures, or test fixes. The clean repro for Bug 1 should make this straightforward to triage.
Thanks,
Steve Mitchell
TopazVideoBugs.zip (63.9 MB)
6 Likes
Very happy to see all the updates recently. That being said Hyperion 2 is completely useless because of the forced .mov output. Hopefully that gets expanded soon. Keep up the great work
Still the same issue. Starlight Precise is not working no matter what video source I use.
logsForSupport.zip (28.5 KB)
1 Like
I went back to V1.4 because V1.5 is not stable on my system, I was getting several times blue screens (I could briefly read something like âStorage Exceptionâ) then the system automatically reboots.
V1.4 and SLP works for me, but limited with input formats issues, so i give Topaz Raw avi files, blows up files but is a workaround. It seems the model is the same, after V1.4 downgrade no new model download happened. I suspect the issue is related to the V1.5 Neuro Server, but thatâs just speculation.
Win11, Ryzen 9900X, 64GB Ram, RTX5090
1 Like
Thank you!! I have been resorting to copying workflow posts to a Word document to keep track of them, an ironically primitive method for such advanced subject matter . There should be a separate Topaz Video category for workflows/best practices posts, to separate them from technical issues and complaints.
3 Likes
Can you tell me if this issue, which I just encountered, is a known problem?
Whatâs the consensus on SLP 4x? SLM 4x was worse than 3x in my opinion, so I never tried 4x, but some people said 4x can give better text output.
SLM has a generall Font issue, also 4x produce Alien fonts, itâs just higher res, smoother, less blocking and because of that it looks better. So if a footage is already fine with no or low compression artefacts, then 2x SLM can be ok, when footage is already blocky, has Hallo effects doing 2x then its not fine structured enough to show small text fonts.
In practice, youâve already lost with small text fonts in videos and SLM anyway. What matters more there is how small pixelated faces are represented, and thatâs where this also becomes relevant. If you only have rough structures because the original is like that and it wasnât upscaled enough, then it simply isnât sufficient smooth too for faces, which is why it ends up looking worser.
But I think this is not really model related 2x 3x 4x it works one same way, it pends on source how high you should go with SLM pending on footage you get it fine structured enough to display something good, with large Lego bricks, thatâs just harder to do.
1 Like
SLPx1 x2 x3 x4
It is important not to upscale to 1080p, as a wide, flashing white line consistently appears at the bottom of videos upscaled to this format. Paradoxically, SLP is not ideal for content below 480p and is also not ideal for upscaling to 1080p.
I donât think Iâve seen that issue when upscaling to x3. Can you even upscale to 1080p? I thought you have to choose x1,x2,x3,x4?
Hmm I havenât noticed it but Iâll check when I test again.
As far as I understand the introduction post, the Topaz Video > General sub category is intended precisely for that purpose. I assume that many users probably only read the release posts, which is why Iâve mentioned it here as well and included links. Itâs also very handy to bookmark the âLatest Postsâ page so you can see all the new posts in every subforum: community.topazlabs.com/latest
For excellent results, you should increase the resolution to 1440p. for example 360px4 or 480px3 or 720px2 even 1080px1 is not ideal.
1 Like
Topaz 1.5.0 crashes every time I try to download models using the model manager.
There are no logs to send since the app just crashes to desktop with no errors.
Feels like there are more bugs/regressions with each new release. Not happy.
In the middle of a render on 1.4 so canât update to 1.5 yet.
Can anyone confirm ifâŠ
1.5 has any speed enhancements?
The estimated remaining time bug on SLP renders has been fixed?
Yes, I had the same issue â even worse: blue screens/reboots during rendering, but it also sometimes happened when no rendering was running, just after playing around for a while with imports and the timeline. I went back to V1.4
1.5 >
No speed enhancements as far as i know. At least thats the feel, i didnât clock it specifically.
Estimate bug is NOT fixed.
did you undervolt/overclock your GPU? Bluescreen and hard reset seems to me is an unstable HW triggered by SW. It wonât happen in 1.4.0 probably because the rendering model has some changes in 1.5.0. Or probably CPU PBO setting as well?
Iâm using it with my 5090 and 9950x3d without issue so far.
