Here’s my personal notes that I’ve taken regarding the published (i.e., live on the public server) models so far:
Non functioning models:
Theia Detail v4
Produces black frames
Theia Detail v6
Produces “Unable to load” error message
Gaia-CG v5
Produces results visually indistinguishable from regular resize (spline, lanczos, etc.)
Models visually indistinguishable from one another:
Gaia-HQ v5 and Gaia-HQ v6
Artemis-MQ v9 and Artemis-HQ v9
Models with missing / incomplete model files on the download server:
Theia Detail v4, v5, and v6:
Missing OpenVINO files, breaks entire libraries
Theia Detail v5:
Missing entire library files for Apple Core ML, breaks entire library
Theia Fidelity v3:
Missing OpenVINO files, breaks entire library
Artemis HQ v6:
Missing Open Neural Network Exchange Single Float (fp32) files, breaks 1x scale mode
Artemis LQ v6:
Missing Apple Core ML files, affects 1x and 4x scale mode, missing one block size file for 1x and one for 4x, library should still function
Notes / differences on other models:
Artemis LQ v6, 7, 8, and 9:
Artemis LQ 6, 7, and 8 are mostly the same. 7 removes a little more noise than 6; 8 is slightly sharper in some areas and less noisy (at the cost of a little detail) than 7.
Artemis LQ 9 appears to be a completely new model, with much improved results. It generates a lot more fine “AI” detail, but also generates some “dither” like noise similar to Gaia CG v6. Liking it a lot!
Artemis HQ v6, 7, 8, and 9:
Artemis HQ is similar to LQ in the sense that versions 6, 7 and 8 seem to be variations of the same model, while 9 is a completely new model. 6 has the weakest effect in terms of focus / denoise. 7 produces much sharper results but also can enhance noise. 8 seems to try to get things even sharper and in focus and also tries to denoise better than 7, but still leaves noise. 9 looks nothing like 6, 7, or 8. It completely tries to denoise the entire frame, with absolutely no grain or “dither” like patterns. It increases focus, but doesn’t tend to generate any artificial detail. Looks less like an “AI” upscale and more like it could have been done with traditional filters, but is easier doing it with VEAI.
Artemis AA 9:
As the description says, it doesn’t upscale, just anti-alias. It looks softer than a regular lanczos resize. Seems to be for special use cases. Not really sure what this mode is doing in a machine learning type of software, because there are plenty of AviSynth / Vapoursynth anti-aliasing filters available already that do a very good job. Not necessarily saying it’s bad, just VEAI doesn’t seem to be the “appropriate” place for this. Just my opinion.
Artemis MQ 9:
At first I thought this was the model from version 1.2.0 that has been re-introduced, but it appears to produce the exact same results as Artemis HQ 9. PNG file size of result is exactly the same. I double checked my JSON files and model folder, it’s definitely loading the MQ files and HQ files for each mode.
Gaia HQ 5 & 6
These versions are visually the same. I haven’t experimented too much yet with them, so I’m not sure how “on par” they are with VEAI version 1.6.1 Gaia-HQ. (Yet.)
Gaia CG 5 & 6
I have so far not found any material that GaiaCG 5 does anything with that looks different than a regular lanczos resize. GaiaCG 6 looks much closer to VEAI 1.6.1 Gaia-CG. In 1.6.1, there seems to be a slightly higher level of what looks like a small amount of post-blurring and post-sharpening (not that it actually does that, just what the result looks like). This can lessen the “dither” types of patterns that it often creates. GaiaCG 6 doesn’t seem to have this “post blur” effect and is capable of creating more subtle, finer detail than Gaia CG in VEAI 1.6.1. But that means that the dither-type pattern is going to be more visible at times.
Theia Fidelity 3 & 4
These two are very close. Need to look close to see visible differences. 3 leaves a little more noise, but can sometimes preserve details which aren’t noise that 4 will remove. The difference is visibly subtle. I would probably go with Artemis HQ/LQ v9 before I went with this. But it can depend of the video, whether it’s animated, etc.
Theia Detail 3 & 5
Theia Detail 4 produces black frames while 6 produces an error message that the model doesn’t load. 3 and 5 look very close visually, but 5 can bring out more detail in small areas. You’ll have to look close to see it, but when you find it in still frame a/b compare, it’s there. Probably not noticeable on video playback. Version 5 is only available with ONNX acceleration as it’s missing crucial model files on the download server for both OpenVINO and Apple Core ML currently.