This backdrop photo does not exist

[Ed Kapuscinski]

I've been doing a deep dive into the uses of generative AI for work.
I've also been struggling to find enough good sources of photos to composite together into a backdrop for the downtown York section of the layout.

Well, combine one with the other and...





Ok, not bad...

One of the cool things about this is that the weakness of generative AI created photos, that they look good from a distance but don't bear scrutiny up close is actually a strength when it comes to backgrounds.

I don't use these images in their entirity, but like all of the photos I use they just provide some content that can be included in a larger composite. Here they are at work.



I gotta say, I'm pretty excited about this development. It's not a cure all for finding good backdrop photos, but it's another really handy tool in the toolbox.

The photos above were generated using Adobe Firefly, which I really like for this, because of the way you can control the prompts, but a coworker tried this out in Mid journey and it seemed even better, so I think I'm going to start playing with that soon.

[Lemosteam]

I need a primer on how to create that giant a$$ hi-res image.  I struggle with this- BLOG?

[Ed Kapuscinski]

I need a primer on how to create that giant a$$ hi-res image.  I struggle with this- BLOG?

Honestly, the only real answer: Spend 25 years dicking around with Photoshop.

I've used almost all of the tricks I've picked up over the years in doing the composite.

It starts with finding the source images, finding the parts you want, and pasting them in. Then using layers to arrange them.
I luck out by going with gray skies, I just make the background white and then bring in the sky at the very end. That makes it easy to use the various image adjustments (curves, levels, etc...) to get things where I want them.

I wish I could explain it all better, but like I said, it's years of practice.

[basementcalling]

I need a primer on how to create that giant a$$ hi-res image.  I struggle with this- BLOG?

I went to the paper mill in West Point, VA, floated out into the river with my kayak, and shot 160+ photos with my Samsung Note 10 phone. I was able to use GIMP, an online photo editor, to put together a 10 foot long backdrop without too much learning curve. Layout never got to the point that I printed the whole file, but I did put a couple small 8 1/2 x 11 prints from Wal Mart self service machines behind a few buildings to check effects.

The learning curve is there on photo editing, but it doesn't have to be as painful as Ed's was. Of course, he can probably do a lot more than I did, as all I was doing was stitching my best shots into one long continuous scene. The magic tool and stitching functions did the hard work.

[Lemosteam]

Its more about starting with a low res and ending up as Hi res.  Does this hacve something to do with RAW?

[basementcalling]

Its more about starting with a low res and ending up as Hi res.  Does this hacve something to do with RAW?

UP3985 is the RAW expert, but it is a file type that allows easier manipulation of the bytes with less compression than if you save as a JPEG file. With today's large megapixel cameras it matters less than it used to I think, but having the capability to shoot in RAW isn't a bad thing.

[timwatson]

You can do a lot with a local version of Stable Diffusion. I think the quality is quite a bit better too. Let me know if I can give you any tips.
I produced a storytelling video of Jesus' birth for YouVersion with that tech. It was fun.
These are my for fun images/illos I've made.  https://instagram.com/wittmason.art

[Ed Kapuscinski]

Its more about starting with a low res and ending up as Hi res.  Does this hacve something to do with RAW?

Ahh, well, the trick there is that you can take a whole bunch of "low" res photos and combine them to a larger one. In my case, on the example I provided, I have at least half a dozen different photos providing the source for what I've got there.

That said, Photoshop is getting scarily good at upscaling images too.

[peteski]

Its more about starting with a low res and ending up as Hi res.  Does this hacve something to do with RAW?

RAW is a bitmap (image) storage format which is basically identical to what the image sensor of the camera captured.  If the camera stores the image as JPG, then the original image is compressed, resulting in loss of quality.  But nowadays, especially for the intended use will be for the image (a model RR layout backdrop), JPG quality should be sufficient. 

As far as the resolution goes, regardless of the format the image is stored in, you can't take low res photo and make it into a clear and sharp high res. photo.  There is not enough detail in the low res. photo for that.  In digital photography resolution usually describes how many pixels are in the photo.  Usually in XY dimensions:  You can take a low res. image (let's say 640x480) and re-sample it to higher resolution (3840x2280), but the resulting high res. image will be blurry.

Then there is the dots-per-inch (dpi) specification for printing the images.  To me printing the backdrop images (viewed from couple feet away) , 200 or 300 dpi should be more than sufficient.  Maybe even 150 dpi. If the backdrop image will be 12" tall, then the vertical dimension of the bitmapped image would need to be 2400 (for 200 dpi) or 3600 (for 300 dpi) pixels.

[Night Owl City]

The photos above were generated using Adobe Firefly, which I really like for this, because of the way you can control the prompts, but a coworker tried this out in Mid journey and it seemed even better, so I think I'm going to start playing with that soon.

You might also want to give Bing image Creator a shot, it is pretty generous on lots of renders, though it requires a free Microsoft account (not that I have not seen any limits on the Firefly beta).

I've been messing around with these for while and with my silly prompts Bing seems to come ahead of Firefly. I haven't paid for Midjourney yet. That seems to be the best for visually pleasing images. I've also recently been using Deep Floyd IF and locally on my Mac, DiffisionBee.

Outpainting seems like it would ideal for generating images to fill the gaps in your composite. DiffusionBee, or another local StableDiffusion, or something like DALL•E let you extend an existing image with prompts. One can only assume a generative version of Content Aware Fill will show up in Photoshop one of these days.

[Night Owl City]

One can only assume a generative version of Content Aware Fill will show up in Photoshop one of these days.

That was quick: https://www.adobe.com/products/photoshop/generative-fill.html

[freescopesdad]

What type of printer investment are we talking about here? Or do you use a commercial or local drop-in shop (Fed-Ex, etc)?

[Ed Kapuscinski]

What type of printer investment are we talking about here? Or do you use a commercial or local drop-in shop (Fed-Ex, etc)?

I paid about $750 for my Epson SC-T2170 through B&H.
https://www.bhphotovideo.com/c/product/1558281-REG/epson_sct2170sr_surecolor_t2170_24_wireless.html/?ap=y&smp=y&lsft=BI%3A514

I'd be about even price wise if I paid for commercial printing for my current layout and now I can do so much more.

[motofavorite]

Ohhh... Why stop at backdrops? Generative AI, a laser cutter, and a 3D printer and never worry again about finding era-appropriate structures to complement those signature structures. And if your training set is 20s-era single-sheath boxcars or wood cabooses, you might get some distinctive but realistic appearing boxcars and cabooses for your freelanced railroad. Merely an automated method of “protolanced” techniques, at an abstract level.

[Ed Kapuscinski]

Ohhh... Why stop at backdrops? Generative AI, a laser cutter, and a 3D printer and never worry again about finding era-appropriate structures to complement those signature structures. And if your training set is 20s-era single-sheath boxcars or wood cabooses, you might get some distinctive but realistic appearing boxcars and cabooses for your freelanced railroad. Merely an automated method of “protolanced” techniques, at an abstract level.

I think that future is coming sooner than you might think.