Hello everyone,
I’m encountering a strange behavior with UniFab in Upscale mode and wanted to share my findings.
My workflow:
Deinterlace
Stabilize
Denoise
The intermediate result after the Denoise step has a High profile with CABAC – technically flawless.
When I then upscale from 1080p to 4K using the Enquoix model (with the best available output settings), the file does get upscaled, but the image quality is worse compared to another workflow.
The better workflow:
If I take the file after Stabilization and process it directly in Denoise mode, setting the output resolution to 4K, I get a significantly better result – without even using the dedicated Upscale mode.
The issue clearly lies in the encoder parameters, not in the AI itself.
The Enquoix model (pure Upscale) forces the encoder into the 20-year-old Baseline mode – a compatibility-only mode without CABAC and without B-frames. This wastes 30–40% more bitrate to achieve the same quality.
The Denoise+4K model (Denoise + Upscale) on the other hand delivers at least Main@L5.1 with CABAC. That’s the bare minimum for modern, efficient HD/4K encoding. As a result, it can show noticeably more detail at 61 Mbps than the pure Upscale mode at just 53 Mbps – despite using only a little more bitrate, because the data is distributed far more intelligently.
From my logs:
Enquoix base model (Upscale only):
Log shows: encoder library used : x264 core 164 (software encoder).
Parameters: rc=cbr and cabac=0 → Constrained Baseline without CABAC.
Denoise+4K model (Denoise + Upscale):
Log shows: CUDA GPU acceleration for video H264 encoding enabled – the NVIDIA hardware encoder (NVENC) is used.
The hardware encoder is extremely fast but usually inferior to software x264 in compression efficiency.
HOWEVER: In my case, NVENC still delivered the Main@L5.1 profile with CABAC – while the software x264 remained stuck on the primitive Constrained Baseline (no CABAC).
With these results, one has to ask: What’s the point of a separate Upscale mode and so many different models, if the integrated Denoise mode (with 4K output) gives better results?
The root cause is clearly the encoder profiles that UniFab enforces in Upscale mode – not the AI models themselves.
Has anyone had similar experiences or found a solution?
Is there any way to manually force the High profile with CABAC in Upscale mode?
Thanks
