Normal and Strong v2 Model Discussion

Happy with V.2

I tried it with selecting the GPU, Auto, and CPU in Preferences. The only one where Normal v2 and Strong v2 Denoise work is CPU.

p.s.
Updated from 531.68 Nvidia driver to today’s hot fix 536.09 without resolving the issue.

Normal V2 smooths out of focus details vegetation in the background too much

Normal v2 is vastly superior.

Hi - I find the normal v2 to aggressive of the faces. I would like the option to pick and keep the current model. Also, I cannot find any way to pick Normal, Normal v2, strong or strong v2 in the autopilot settings. So right now, I cannot use V1.3.7 because the normal v2 makes sports images look unnatural.

Love the new V2 model.

I think they’ve discovered the cause of the problem for certain GPUs.

Did you try downgrading to Studio v528.49?

No, I stayed on the newest release, and they have it working for me now. They’ll probably want to run a few more tests before they make a public release of the model updates that were needed.

I just used the Strong v. 2 model in PAI 1.3.7 - Win 11. Standalone.

I liked the way it retained a slight bit of noise and didn’t over-plasticize the skin on the woman in my image when I kept the Strength slider low (@ 8). Ditto for the bokeh foliage background, it left enough noise so that it wasn’t over-smoothed.

Lower the value of the denoise slider.

In both models if a person has glasses on, the frames of the glasses are too sharp compared to the rest of the face. (Both for normal and normal v2). (Normal v2 works better for very noisy images compared to normal).

Denoise Strong_v1 vs Strong_v2 (ISO12800):

Denoise Strong_v1 vs Strong_v2 (ISO12800):

Denoise Strong_v1 vs Strong_v2 (ISO12800):

Denoise Strong_v1 vs Strong_v2 (ISO12800):

For fine grain noise, the Strong_v2 give cleaner & sharper result.

Download the original image from dpreview

1 Like

Both strength and detail only set to 10, however it was set to upscale when I opened the image, setting that to 1x made a difference - testing further.

However Denoise Strong_v2 doesn’t work well with image that contain color spotches.

Here is a example,

The Left Image is processed by Denoise AI v2.0 from 2020 using Clear AI.
The Right Image is processed by TPAI V1.37 Denosie Strong_v2 (Strength 100, Detail 1).

Denosie Strong_v2 tends to left all the color spotches in the image, while clear AI able to remove it.

Download original image (cropped)


Here is another example,

The Left Image is processed by TPAI V1.37 Denosie Strong_v1 (Auto).
The Right Image is processed by TPAI V1.37 Denosie Strong_v2 (Strength 100, Detail 1).

Download original image (cropped)

Normal v2 is much worse than normal. leaves a lot of noise area like staines. especially on light edges of subjects. Also has the same plastic look as strong filter that i never use. because it makes subjects look artificial or cartoon like

I think this is because the image has already been traditionally denoised.

Either from the camera or in the RAW process.

Strong V2 tries to keep the colors as they can be important information.

And to do this with a finished image is difficult, with a RAW is easier, so Adobe and DxO with their RAW denoisers still have an advantage but only if you can go with a raw in the process.

During the beta it turned out that both new models can have problems with high ISO (ISO 6400+) images, so the old Strong model was added again.

Hi, could you show an example?

The image is straight out of camera without any post-processing afterward.
Old Sony DSLR/SLT camera tends to leave color spotches in image at high ISO.

Here is a crop from the SOOC image from camera.

When its a jpeg (SOOC) and not a raw it has been processed by the camera.