My god, these are coming out QUICK, as I just DL’d un/installed the previous one, booted it up, and bam, another beta! I’m so glad I invested my time and money on this product!
EDIT: I’m guessing this is an isolated feature that we’re testing, and it won’t be permanently cut to only 4x so no worries here!
Is there a way to have the “instances” variable set to 0 instead of 1? It’s causing processing failure on some models, which wasn’t happening in v3.2.1.
For me, model selection works, I processed that destroyed D&D cartoon through the preview on all 4 of them, and, unsurprisingly, v3 is the fastest. All Proteus models downloaded correctly and got used. And it seems they’re converted to TensorRT format (in my user case and with exception of v4)!
I’m noticing “starry” artefacts when Apollo Fast v1 model is being used, but that’s probably because it’s in beta. It usually shows on scene changes or more dynamic movement interpolation - which is great, I have to admit. It even (finally!) corrects the movement irregularities almost perfectly in that “Lalafell Reaper” video I’m testing VEAI’s motion interpolation on regularly.
The new Proteus v4 checkpoint, when set to high Revert Compression values, crushes blacks or makes darker colors go black. On lower values it’s less pronounced.
Can confirm with this bug. Can only try V4 when that previous model selection tick box is unselected.
In relation to V4, obviously work in progress:
The slider bars still do next to nothing. They do have an effect but not much of one.
The noise reduction is extreme by default. Even at Manual with 0 noise reduction, the output is heavily noise reduced. Unfortunately it makes the output lose too much detail and as the slider is already at 0, you can’t adjust it.
There needs to be a slider bar for the new face enhancement. Faces in background are often better, but it also tries to sometimes enhance a face that is blurred intentionally. That is, when the camera is focused on a foreground person and the background is thematically blurred (not just bad quality) IE here the below image. You can see she is supposed to be blurred as background by the AI has tried to sharply enhance the face leading to a rather creepy result. Interestingly, in the further down image the character “in focus” has had far less focused enhancement done so actually looks (weirdly) more blurred than the background character.
Artemis has generally speaking always retained more detail than Proteus. I still use V8 of Artemis for the bulk of upscaling for that reason, as V9 and later softened things too much.
Proteus has always been more of a glorified Avisynth script that allows for “more friendly” processing of common issues - wanting to sharpen a bit more, wanting to denoise a bit more etc - but lacking much of the AI that seems to be in the upscaling as no matter the countless hours I have spent with Proteus 1, 2 and 3, none can replicate the actual level of detail expected from upscaling with AI.
Note that the Proteus Auto settings for V1, 2 and 3 have always erred on the blurred side. Proteus doesn’t work unless you dial in the knobs as it were. I am not sure from the comparison images if it was just “Auto” for 1 and 3, but guaranteed the output of Auto on those models won’t hold a candle to Artemis - on virtually any version.
Its why I am excited for V4. The face reconstruction is the first actual AI progress I have seen in Proteus since it was first released. But even on V4 with sliders mostly broken, I have to set Detail to 100x Sharpness and Deblur much higher than you would put on V3 or V1 just to get any effect, which is better.
The below comparison Album has options for Proteus V4 Auto, Proteus V4 Manual, Original and Artemis LQ V8. Note that the Artemis one had more denoising done before feeding into the upscaler as it amplifies noise:
If the detail retention of Artemis can be added to Proteus, with a slider for Face Enhancement and fixing the slider bars so it doesn’t denoise out the wazoo, it may be a perfect amalgamation.
I think VEAI also tries to correct whatever kind of color tint compression artifacts can create. I like it. Though, let’s hope it’s not “hardcoded” into the models, so, in the future, it could be optional - more choice to have is (almost) never bad!
I have to say, the quality of Proteus v4 is much better! Good work! With manual settings, that is.
However, the ‘Dehalo’ function is currently too exaggerated, even at a value of zero.
With more testing, I find that ‘Revert compression’ ‘Reduce noise’ and ‘Dehalo’ are set too high and even with settings of zero it’s too much sometimes, as it removes subtle details (such as hair or skin texture).
