Using AI to create and edit textures.

[EldoradoRailroad]

Should have asked the LLM to put glasses on Jack Benny!

[wwhall]

Here is more of what I've discovered with GG. What GG doesn't "understand", it will often "bork" because it is only looking at pixels. So, if it does not understand a "detail" in a photo or screenshot, it will "hallucinate" something. One way to prevent this in, say, a screenshot, is to use the word "exactly" in the prompt--for example, "render a photorealistic image of this exact locomotive in the scene, while making the surrounding scenery realistic." I've found that using word "exact" tends to force GG not to modify the appearance of the locomotive, though it may change its perspective. Also, since GG is not modifying the appearance of the locomotive, it won't throw a hissy fit about logos or trademarks. The best way to keep GG from giving the "I can't do that" message because of a trademark is to also never refer to something by its trademark name. For example, saying "General Electric" or "Union Pacific" in the prompt may get you the "I can't do that" message. What GG will allow you to do is to exclude brand names, trademarks, etc. to narrow it to what you want. For example, "Please render a photorealistic image of a modern American passenger train being led by two passenger locomotives, neither of which built by EMD or Siemens" to force it to render a GE locomotive. I know, it sounds stupid, but it will often work. ALWAYS try to write the prompt to get what you want on GG's first render. If you ask GG to modify it in a subsequent render, it will usually start borking or hallucinating stuff that you didn't want it to do. If you do have to make an additional prompt, clearly state what you want changed, and then add, "Change nothing else in the scene." Unfortunately, even with that prompt, GG may still bork stuff, but it helps. GG does neat stuff, but it is really pretty stupid. And, don't be afraid to call it out when it screws up. One day, I asked GG to render an image of a city hall at night in a particular city. It rendered a "generic" city hall that barely resembled what I wanted. So, I uploaded a daytime photo of that particular city hall and told GG "this is the actual city hall that I wanted rendered in night scene." GG apologized profusely, then rendered that scene that I wanted.

[ossie]

There is actually some humour to be found in it at times amongst the frustration isn't there Wade? I don't know if it's part of their attempt to crack down on deep fakes/trademark infringements etc but of late I've regularly had "sessions" with it where it will go through the usual "considering this", "rendering that", "checking the instructions" etc and then instead of spitting out a photo I get a wall of text describing the photo. After 3 or 4 of these in succession recently I said something along the lines of "can you just stop talking about it and just show me the damn picture I asked for?". It came back with "I can tell from your tone you are displeased with me and I profusely apologise. It appears I have over-promised on my ability to supply the picture. Something (it gave a technical term) is missing from my toolbox". I wasn't sure if I was supposed to feel sorry for it for chastising it or whether that was AI speak for "shove it mate, I'm doing my best".

[oper190]

I've been using paid version of GPT and the free version of GG. Both have their limits, but so far GPT is more open minded, and will explain in better detail what is wrong with your prompt, and offer revised versions that it can do. GG just say things like "that violates my rules" without telling you what part of the prompt is the problem. Again I don't know if a paid version of GG would be more open minded, or helpful.
I've been giving both GPT and GG the same exact prompts, to see how each will render the same scenes, and have found GPT rendering without question prompts GG won't do.
Then there are just little things that put you over the line. Example, gave GPT five images of people from photographs, with one a general view of the pool area at a resort I was at to use as the background, asking GPT to put these people in a scene. I stated using the photos as 'references for face, hair, and body shape and form' wearing swimsuits, one in just shorts and a t-shirt with a camera taking a picture of the others. GPT tells me it can't reproduce real people like that ( which I've done many times with that same prompt ). However GPT tells me it's the use of camera, and taking a picture of the other people in revealing clothes that's crossed the line. So I modified the prompt, removing the reference of a camera, and it rendered the scene, or I could have to stated make "fictional" people, meaning they'd never look anything like the people in the photos, which usually do get morphed a little anyway. At least GPT offers alternate prompts it can do, or explains the "why" it can't do what you asked.
GPT does give me outlines of my prompts sometimes for clarity, so I copied one to notepad, which I now use as a basis for most prompts, and am getting less push back, and better explanations if it can't do them. - Can't say what a paid version of GG would do. - Also with the outline it makes me think out the scene in more detail then I was giving before.

[oper190]

Here's a sample of the prompt I was given:

Scene Prompt:
Create a photorealistic scene with the two women and man pictured as a reference for face, hair, body shape and form. ( Describe the basic scene wanted )

Woman 1 ( dark top ) describe what she is doing and what you want her to wear

Woman 2 ( pink top ) describe what she is doing and what you want her to wear

Man 1 ( gray suite ) describe what he is doing and what you want him to wear

Environment details:
A well lit outside daytime scene in a park......
( include things like well lit, dimly lit, what is going on in the background, with a blurred or sharp background, windows or not, colors, personal effects...etc )

Camera style:
High-quality DSLR look, sharp focus, natural skin texture, realistic lighting, true-to-life proportions, cinematic scene, focus is on the three people

[wwhall]

I've been using paid version of GPT and the free version of GG. Both have their limits, but so far GPT is more open minded, and will explain in better detail what is wrong with your prompt, and offer revised versions that it can do. GG just say things like "that violates my rules" without telling you what part of the prompt is the problem. Again I don't know if a paid version of GG would be more open minded, or helpful.
I've been giving both GPT and GG the same exact prompts, to see how each will render the same scenes, and have found GPT rendering without question prompts GG won't do.
Then there are just little things that put you over the line. Example, gave GPT five images of people from photographs, with one a general view of the pool area at a resort I was at to use as the background, asking GPT to put these people in a scene. I stated using the photos as 'references for face, hair, and body shape and form' wearing swimsuits, one in just shorts and a t-shirt with a camera taking a picture of the others. GPT tells me it can't reproduce real people like that ( which I've done many times with that same prompt ). However GPT tells me it's the use of camera, and taking a picture of the other people in revealing clothes that's crossed the line. So I modified the prompt, removing the reference of a camera, and it rendered the scene, or I could have to stated make "fictional" people, meaning they'd never look anything like the people in the photos, which usually do get morphed a little anyway. At least GPT offers alternate prompts it can do, or explains the "why" it can't do what you asked.
GPT does give me outlines of my prompts sometimes for clarity, so I copied one to notepad, which I now use as a basis for most prompts, and am getting less push back, and better explanations if it can't do them. - Can't say what a paid version of GG would do. - Also with the outline it makes me think out the scene in more detail then I was giving before.

Next time that GG does that, do this: Ask GG why it rejected the request and would not render the image and what could be changed in the prompt to avoid the rejection. GG will usually tell you. You may not like the answer, but it will give the specific reason why it was rejected, and it will even sometimes "suggest" a workaround. What I learned from doing this is that GG has a "two-sided" brain. One side works on the technical aspects of rendering the image--once that "side" figures out how to render the image, then the "other side" decides if the request violates any of the GG "rules." I have had the "censor" side give a "rejection" message, but the "technical" side has gone ahead and rendered it, anyway. Just for information, GG has gotten "overloaded" several times (most recently a couple of days ago) and quit rendering images at full resolution. When I asked GG why, it apologized and admitted that the system was being overloaded and that "they" were working on expanding its capacity.

PS--For GG, I've learned never to refer to an uploaded image as a "photo"--that triggers alarm bells in GG because GG insists that it is NOT a photo editing program. Always say something like, "Please use the uploaded IMAGE of a locomotive to render a photorealistic scene showing the locomotive sitting in front of a railroad depot in a mountain location in winter at night."

[oper190]

PS--For GG, I've learned never to refer to an uploaded image as a "photo"--that triggers alarm bells in GG because GG insists that it is NOT a photo editing program. Always say something like, "Please use the uploaded IMAGE of a locomotive to render a photorealistic scene showing the locomotive sitting in front of a railroad depot in a mountain location in winter at night."

I always refer to either Screenshot or Image for my references
From GPT I've learned also using any terms like: pose or posing, taking a picture, implying taking picture, direct intimacy ( can be implied by facial expressions leaning into each other are okay) but no kissing and cuddling. So long as not posing or implied to be a picture, a two-piece swimsuit, full-slips for woman are okay, but a bra and half-slip are out.

[ossie]

Again I don't know if a paid version of GG would be more open minded, or helpful... Can't say what a paid version of GG would do..

I've got the paid version of GG Geary and basically the only difference between it and the free version that I have found is that the regular "what the?" moments now come with a price tag. There's no doubt that their understandable attempts to slow up deep fakes has had a much wider implication for users and their images than you should probably expect from an intelligent system and even allowing for it being early days, there are still things happening way too often that drive you crazy. It is key to learn to "speak AI" as far as instructions go, but even then, the number of times those instructions are ignored by a system that is supposed to get better as you use it does bring about some frustration.

One area in particular that has pushed my buttons considerably is image resizing. Variations of "don't change the original image size" were getting the same level of attention as "don't eat any more lollies before dinner", particularly as height may have been left unchanged while width would be shortened, therefore mangling your image perspective. After having a conversation about it with it, (I'm paraphrasing) it told me that it tends to prefer certain "standard" image sizes and wants to work with them as default and the work around for this is clear instructions defining your image size and requesting those dimensions be adhered to (difficult if it is a size that is already larger than their standard output) or using a phrase such as "strict aspect ratio must be adhered to".

Strict aspect ratio must be adhered to is a tag I now add by default to every image request and still I'd say 1 in 6 are being ignored and "squashed" (along with other regular borks) so I really appreciate your discussion about GG & GPT. I didn't spend a lot of time with GPT before I put my eggs in the GG basket so I will go back and have a more in-depth look at GPT, thanks!

[oper190]

I didn't spend a lot of time with GPT before I put my eggs in the GG basket so I will go back and have a more in-depth look at GPT, thanks!

With GPT I am finding it's standards are a little more open, however in my prompting both for the same renders it is basically a coin flip as to which renders better or more accurate to my prompts, on the overall I do favor GPT, though GG tends to render faster.
Had I been able to get one of my credit cards to work for purchasing GG I'd have bought that, so I went with GPT, both had first month free deals at the time.

[ossie]

Oh.... while I think of it, after Chris introduced this thread here, I took it over to World of Tanks and a bunch of us have had similar fun with that game as we are here in the sim. Likewise the discussion has turned to AI tools but one of the guys in our clan is a BIG user of GIMP and maintains you don't need AI if you know how to use GIMP properly. I've never tried GIMP for the sim but I know plenty of others have. If any of you happen to read this, is that a fair assessment? I'm not so much interested in the turning screenshots to photo aspect, but deep image modification such as realistic shadow removal, upsizing images without losing quality, the ability to work on lens blur, distortion etc without destroying small lettering and numbers (aspects that Photoshop doesn't do well and I'm hoping for AI to handle), are they things GIMP will do effectively?

[superheatedsteam]

are they things GIMP will do effectively?

GIMP isn't going to do any better than Photoshop. I don't use Photoshop but doesn't it already have AI tools to assist in image manipulation?

You may want to look at running an AI system on your home computer as you can customise it for the specific workflow that you want. ComfyUI is not the only option but appears popular. The following youtube link (of which there are many) demonstrates how to install and run it, how to download 'upscale' plugins and then upscale an image. There are other plugins for other specific areas of image manipulation that you download to customise your specific workflow.

https://www.youtube.com/watch?v=g74Cq9Ip2ik

And its all free.

Cheers,

Marek.

[ossie]

Thanks for the info about GIMP Marek. I thought it was worth asking the question here as if it was as good as my WoT mate said, I assumed trainsim people would have been shouting from the rooftops about it as THE editor of choice for years.

Topaz Photo AI does work with Photoshop as a plugin but while it has been terrific for upscaling and cleaning up photos and slides I have not been able to hit on a filter/setting that doesn't damage the sort of small text we're dealing with for train related pics. Just with the skimming I've done through AI, Reddit etc, it seems to be a common complaint and the prevailing reason/theory seems to be that it considers small text details as "noise" and tries to replace the pixels as part of its regular upscaling method. There's a mountain of Topaz tutorials on YT so I won't completely throw in the towel until I've had a look through them to see if there's a solution.

Thank you for the link to ComfyUI. In trying to find a solution to the Topaz text issue I did see a number of people suggesting DIY AI setups that could be "trained" to your personal specifications. As I started digging into AI editing I was really surprised to find that there isn't an all in one type package that allows the level of interaction that GG & GPT have vs the much better imaging results that Topaz has given me. (As an aside, I did try one called Luminar Neo that allows the chat function but that is still limited to discussion of using the features the program has. It won't go and find a replacement for an area of shadows for instance the way GG would). I'm sure that an all in one package will arrive at some point but the price tag will be interesting. For instance, Luminar Neo offered me a perpetual license for about $130 which seemed VERY reasonable until I saw the fine print saying that "large" updates would incur additional costs as they were made available.

For a lot of what I'd like to do with photos, slides, "PC art" etc, the self customisation route might turn out to be the better option for me, so again, thanks very much for the link. I appreciate it.


NOTE: In fairness to Topaz and anyone that is considering trying any of these packages, my complaints probably should come with a couple of caveats. Firstly, in relation to this hobby, I've done very little so far with trainsim related images and only a couple of TS related "tests" have gone as far as being tried in the sim. The level of "damage" I'm complaining about on a 6k image in Photoshop may not be a big issue by the time it's reduced to a 2048 image on moving models for ORTS. Just on the couple of tests I've done, I have no doubt that the results I'm complaining about are still going to be better than 512 or 1024 images, so please don't let the above comments put you off trying one of these packages if you're looking for upscaling abilities. The flipside of my complaining is that if a smaller and/or blurry image can be increased in size/cleaned up to make it a viable texture, even if some of the small detail is borked, we're still better off aren't we?