I have some old 110 film strips that were scanned by a professional scanning service. The camera used was an old crap Instamatic from the 1980s. The combination of a lousy camera plus 110 size film plus scanning means that the digital versions I have are very blurry.
I know there is a limit to how much can be done, even with the generative models, but I am having trouble finding a decent model in Topaz. I’ve tried Wonder 3, Low-res, Recovery 2 (with predownscaling), and Recovery 3.
Recovery 2 is the only one that makes a substantial difference; the others improve the blurriness very slightly, but not enough to consider the result acceptable. The problem, of course, with Recovery 2 is that it produces noticeable artifacts (such as smeary areas), and faces are so clear that they look artificial.
Are there any ideas out there about how to get an acceptable result?
I don’t know why it’s not working for you but, you could try Copy & Paste or converting your files to PNG, or JPEG alternatively, when Topaz Support contact you they’ll give you a Dropbox Link
It’s not perfect, and the faces might be changed a little bit, but considering the lack of details on the subject I think it’s good. I wasn’t able to get the glasses on the woman on the right, the AI removes them
I agree that your efforts produced a really good result given the poor quality of the original image!
Why did you choose Redefine-Realistic vs Recovery?
One of the things that I find difficult is to choose among the large number of models that topaz provides. Whether I use local processing (I have an M4 Studio Max) or cloud processing, it takes too long to try multiple models and multiple settings with each model in order to see which combination seems to produce the best result. While there are some general guideline (e.g. it would be stupid to use High Fidelity on this image!), for any given image there seem to be multiple possibilities that are reasonable.
Typically with scanning, over-scanning will occur in an effort to make a larger digital image; but this often introduces false resolution. In all of the examples below, some form of downscaling was applied to condense image detail before or between processes.
I usually manually downscale whenever I want to condense image detail.
1x of Standard Max, Recover 2, or Redefine realistic can help act as a noise reduction pass and a clean up for dust and scratches. But each can tweak the faces in undesireable ways, so I tend to enable Face Recovery if needed.
Wonder 2 or 3 is the best final upscale step for accuracy. If you’re comfortable with compositing, doing an additional process with Redefine realistic Subtle or Redefine creative Low will give nice environment and background details you can brush in with layers in your editor of choice.
As you can see, different people have slightly different workflow and results. Our models are not universal for a specific task, and this is why it’s important to try a few different ones on restoration like this. With cloud rendering, I usually send the file with all the generative models simultaneously. I can then later compare the different outputs, and if needed tweak the settings on that specific model and send again to the cloud later.
I chose Redefine Realistic in this case as it gave me the best results for the clothing, and for the faces. It recreated the textures accurately. My next step on this image would be to blur the background and de-saturate slightly to hide the issues there. We want the focus to be on the people.
Thanks for the additional detailed information. I noticed that you don’t typically downscale as a first step, but rather apply something else such as Standard Max and Recover Faces BEFORE downscaling. Why?
And, when a model has down scaling built (such as Recover 2), why do you prefer Manuel down scaling?
I’m asking these questions because I want to understand some of the theory behind what you were doing so I don’t feel as if I’m just shooting blindly in the dark to see what steps added, which order to apply.
One of the things I find interesting is how unpredictable many of these models are in whether they alter a person’s appearance too much. By “too much” I’m not talking about faithfulness to the original image (which can be hard to discern if the quality is poor) but looking at whether the end result just doesn’t seem to look enough like the person I know in the image.
Decades ago I saw an exhibit in a science museum that showed how incredibly little information is necessary for a human being to recognize the face of someone they know well. The exhibit took a picture of Abraham Lincoln a repeatedly downscaled the image, and at each step then enlarged it to be the same size as the original. Of course, as the amount of downscaling increased, the blown up image became more and more pixelated. Even when 90% of the information in the original image was lost, I had no trouble whatsoever recognizing the face as Lincoln’s. Apparently, psychologists have studied this phenomenon extensively, and in our species, this ability seems to be limited to faces, not to other physical characteristics.
I failed to clarify that the process with Recover 2 included pre-downscaling set to Low. So that workflow had two passes of downscaling.
Only the 1x Standard Max workflows didn’t start with a downscaling. I find that Standard Max already applies quite a bit of smoothing, denoising, and compression clean up.
Then I downscale before I Wonder, allowing Wonder to generate realistic textures and details. If facial accuracy wasn’t a factor, I would probably finish with Redefine realistic: Subtle instead.
Ultimately, downscaling = condensing details. If you’re dealing with softness, noise, false resolution, artifacts, dust, scratches, etc; inserting a downscaling process can have benefits.
Depending on the source resolution and the degree of the issues I am solving for, I’ll reduce to image widths of 2048, 1024, or 512. Then send all of these with a 4x using Wonder 3 to gauge which downscale is best to work with. Unlimited Cloud render makes versioning pretty quick and easy.
And sometimes I’ll even downscale in Gigapixel using Standard, Low res, or High Fidelity as the downscaling model. I’ll usually try scale factors of around ~.7x, .5x, and .3x.
Every image is different. Having a supply of techniques at your disposal is key.
