"Draft quality" mode for AI raw denoise

[743coffee]

Hello,

I have found the AI raw denoise module to be very effective in denoising images, but it is understandably quite slow on my M2 MacBook (specifically the large model size). This has been a noticeable bottleneck when editing a series of noisy images.

I am wondering if AI raw denoise could benefit from a "draft quality" mode or "high-quality export" mode, where the model would run at quarter size when editing, but switch to large size when exporting. This would make real-time editing much more feasible on non-GPU systems, without having to compromise on image quality. (Having to wait 2–3 minutes any time I want to revisit an AI denoised image can be rather frustrating.)

AI denoise also appears to run on thumbnails, which makes copying AI denoise settings across images quite a costly operation. Draft mode could alleviate this issue as well; or AI denoise could be bypassed entirely for thumbnails (at least when zoomed out).

Thank you for your consideration!

[Annabel2]

Something I haven't tried but which might help here, is to use the Style feature when exporting. So edit the image without denoise for speed, then when exporting, choose the Style you've already set up which just adds the denoise. Then get some coffee while it exports!

[743coffee]

Unfortunately, this will override any image-specific adjustments made to the module, so it is not really a practical solution unless you use the same denoise settings for all of your images.

[pedrorrodriguez]

Not sure if this would work, but you could have a preset for AI raw denoise so that it is automatically applied at low quality (or added but disabled) to every new image. Then apply a style before exporting that toggles the module to high quality.

However I'm not sure if this is so susceptible to automation, since you may want to scale back the amount of denoising on a per image basis.

[aurelienpierre]

AI raw denoise runs on non-demosaiced content, the Bayer and XTrans CFA patterns are part of its input and training. Running it on half-sized or quarter-sized images not only skews the noise statistics, but also implies we do some sort of basic demosaicing. The module can't deal with such input.