Tried Denoise/Enhance on some older Fuji RAF files, and the "wormification" was terrible. I should try again with sharpening down to 0 first and see if that helps.
RoamingScott wrote:
Tried Denoise/Enhance on some older Fuji RAF files, and the "wormification" was terrible. I should try again with sharpening down to 0 first and see if that helps.
I never used Fuji, but I can say that I've gone back as far as 2003 to old Nikon D2H NEF files and Canon 1Ds Mark II files and it really does a great job. Across the board though I like the results with sharpening turned to zero first and then selective sharpening after denoise.
Just tested by starting with 0 sharpness on a RAF and the resulting image is much nicer than having the default sharpening applied at the start. Good to know. Also, lord I hate Fuji RAF files
jhapeman wrote:
I don't think it's a bug at all. They state that the DeNoise includes their "Raw Details" algorithm, which is essentially a type of sharpening baked into the Bayer (or Fuji X-Trans) filter demosaicing process. Because it can also carry over any adjustments you've already done to the file, if you pre-sharpen it will leave that in place but it does appear to apply the sharpening to the file before doing the other steps. I think that's intentional on their part, although I can see it being confusing in some way.
Personally I find that I have now changed my workflow--I am turning off the default sharpening as part of my import preset, then apply the AI Denoise, and then using masks to only selectively sharpen the image. Vastly superior results that the recent mask and DeNoise updates have now opened up!...Show more →
It very well may be intentional. The only reason I brought this up is it is not explained very well. There are users that may not be into these things as much. When you import the sharpening default is 40. If they send it to Denoise they may push further than required. Past 75 or so it does start to look plasticky. You can see in my example how much different the file looks in Denoise with sharpening at 0 and 50.
How do I use it?
Using Denoise is easy:
Click the new “Denoise” button in the Detail panel to bring up the Enhance dialog.
Adjust the Amount slider to taste.
Press the Enhance button.
RoamingScott wrote:
Just tested by starting with 0 sharpness on a RAF and the resulting image is much nicer than having the default sharpening applied at the start. Good to know. Also, lord I hate Fuji RAF files
Bruce n Philly - I've got an AMD Radeon RX 6650 XT (8gb) graphics card on my AMD machine (running a Ryzen 5 3600 CPU). My Canon R5 RAW files are taking 40-45 seconds to de-noise.
PintPlease wrote:
Bruce n Philly - I've got an AMD Radeon RX 6650 XT (8gb) graphics card on my AMD machine (running a Ryzen 5 3600 CPU). My Canon R5 RAW files are taking 40-45 seconds to de-noise.
My 2019 Mac Pro has a Radeon Pro 6800X Duo in it, but it still takes about 27 seconds an 61mpx A7RIV file. Not quite as fast as the Mac Studio Ultra.
Bruce n Philly wrote:
Thanx ... I could not find any information on Radeon cards and LR Denoise AI.
Not sure what I will do... if anything... yet. Big price tag for what? I don't see the benefit of LR over Topaz as I am paid up on my Topaz.
Hmm.....
Peace
Bruce in Philly
Several years ago I learned the hard way with Topaz apps that AI eats VRAM. When I got a new iMac I made sure I had enough. It still takes 45 seconds but so did Photo AO and PureRaw 3. My lowly MacBook Air with the M1 chip is faster.
Trying it out now. Whether I use DxO PureRaw or LR, they both take about 35 seconds per image (20mp Canon R6). I have not tried Topaz so I can't say whether it is faster or slower.
PC is a few years old. AMD Ryzen 7, RTX 3050 12gb, 32gb DDR4 Ram.
jaygould wrote:
Trying it out now. Whether I use DxO PureRaw or LR, they both take about 35 seconds per image (20mp Canon R6). I have not tried Topaz so I can't say whether it is faster or slower.
PC is a few years old. AMD Ryzen 7, RTX 3050 12gb, 32gb DDR4 Ram.
Photo AI will likely be about the same. There standalone apps about half
An argument for waiting to buy a new video card... something to ponder.
I suspect there is a huge population of LR users with older generation video cards such as me with a GTX card (or older) and are having performance issues with denoise AI. The RTX generation contained innovations that LR denoise AI uses for performance.
So, what do I and others do? Purchase an RTX card or wait?
The cheapest RTX card I can find is the RTX 2060 for $349. Correct? Not saying this is best card for denoise AI, but a data point.
Issues IMO:
1 - Denoise AI is pretty good but so is Topaz Denoise AI... for me both are equal in results
2 - Given Adobes history, I suspect they will increase the performance of their new feature over the next say 6 months
3 - I am already paid up on my Topaz support for the next 10 months
4 - Topaz is massively faster than LR and LR Denoise AI is pretty much unusable for me given my older card.
5 - This article from Tom's Hardware notes that new releases from Nvidia and market changes (bit coin collapse, used cards flooding the market, prices coming down) should play into my decision to purchase a new card... waiting will get you a much more powerful card for your money.
6 - There is no other application I use compels me to upgrade my card
So.... I am going to wait to buy a new card and continue to use Topaz.
Those are what Adobe has suggested as I mentioned before. I do wonder now if the AI acceleratio has changed the weighting. For example would a 40 series card that has the same compute results as a 30 series actually be faster in PS/LR AI?
Bruce n Philly wrote:
If anyone is considering buying a new card, I find the Passmark benchmarks to be very valuable... you can kinda find value sweet spots.
Ehh, synthetic benchmarks like that are pretty useless when buying for specific tasks. It's better to focus on performance benchmarks for the actual tasks you intend to use a card for. When it comes to photo editing, you can check out the benchmarks for Photoshop and Lightroom that Puget does on the PC side, although now that Apple Silicon is out Puget has yet to update the benchmarks (IMO because they have zero incentive to do so and the results wouldn't make their high-end PC offerings look all that great in terms of value, but I digress).
Has no website run comparison tests with different nVidia RTX cards in PS/LR to determine the relative performance in "AI" accelerated functions? It should not be too difficult.
Anyone found a reliable site listing the number of Tensor cores on different RTX cards? even the Nvidia website doesn't seem to state anything but generation.
mcbroomf wrote:
I did it manually, I was going to add all the other attributes but this is as far as I've got
GPU Tensor cores
2060 240
2060 Super 272
2070 288
2070 Super 320
2080 368
2080 Super 384
2080 Ti 544
3060 112
3060 Ti 152
3070 184
3070 Ti 192
3080 280
3080 Ti 192
3090 328
3090 Ti 336
4070 240
4070 Ti 184
4080 304
4080 Ti 440
4090 512
4090 Ti 568
PS. Please feel free to 2x check and let me know if I have a typo or transferred the wrong number somewhere
I wonder how the Tensor Cores compare between card generations. I would think the 240 Tensor Cores in the 4070 would be faster than the 240 Tensor Cores in the 2060.
