Heads up: Since I wrote this I have made the StretchScape model which is an improvement upon the method in this article.
https://civitai.com/models/99310/stretchscape
Intro
To be honest I didn't do too much testing to see how much I could push the aspect ratio with this model when I was making it. I was testing upto around 16:9 and I was really happy with the way it was handling this. However upon releasing the model, user AugmentedRealityCat made a comment about how well the model was generating the wider landscapes prompts and settings he was passing it.
So this has been rattling about in my head for a few days whilst I've been at work. I have seen the results of aspect ratios upto 3:1 as I was using this for some page headers and it worked great. For my Ko-fi header I just literally typed in the needed dimensions and used the image as it turned out, nothing else.
So before I try an aspect ratio of 4:1, I thought I'd just post the results as I get them from Auto1111.
Feckit, go bigger - 5:1 if we don't explode in a BOOM of pixels and data before we get there.
Test
As we have 10 slots for up-loadable images I think we will go for...
2 * 1:1
2 * 2:1
2 * 3:1
2 * 4:1
2 * 5:1 (we will work something out with the resolution if we get that far without loads of repeats in the 4:1 image)
So 2 different prompts with same settings for each aspect ratio. I have a thing for small cute robot droids (R2D2 with arms and legs basically) made from discarded junk, so that will be one prompt. For the other I also like just typing in fantasy dream landscape or something to see what the model produces. So I'm going to hit ChatGPT4 over on nat.dev (honestly, check it out. All the latest models on pay as you go instead of monthly subscription. GPT4 knows what a SDprompt is off the bat and churns out some sweet ones).
Prompts from GPT4:
Stoke:
Can I have two detailed StableDiffusion prompts with appropriate negative prompts please?
Prompt 1 involves a cute robot droid made out of household junk watching the stars in a detailed realistic widescreen image (be as expressive as you want as i would like a really stunning image please)
Prompt 2 i would like the most beautiful imaginative fantasy landscape you can really impress me with. Image must be detailed and visually stunning and eye catching.
GPT-4
Prompt 1:
A captivating image of a cute robot droid, skilfully crafted from household junk, standing atop a hill in a serene, moonlit landscape, gazing up at a breathtakingly detailed star-filled sky, captured in ultra-high-definition 8K resolution, utilising realistic lighting and atmospheric effects, inspired by the evocative visual styles of artists like Hayao Miyazaki and Simon Stålenhag, showcasing the wonder and curiosity of this adorable little being in a stunning, awe-inspiring scene.
Negative prompt: blurry, low resolution, poorly rendered, lack of detail, flat colors, unrealistic lighting, no stars, cluttered composition, unexpressive robot, cropped widescreen.
Prompt 2:
An incredibly beautiful and imaginative fantasy landscape, featuring a mesmerising blend of floating islands, cascading waterfalls, enchanting forests, and otherworldly flora and fauna, all bathed in the warm, golden light of a setting sun, captured in ultra-high-definition 8K resolution, utilising vibrant colours, intricate details, and breathtaking visual effects, inspired by the awe-inspiring visual styles of artists like Roger Dean and James Gurney, presenting an eye-catching and visually stunning masterpiece that transports the viewer to a realm of wonder and magic.
Negative prompt: dull colours, lack of detail, generic landscape, flat lighting, uninteresting composition, low resolution, poorly rendered elements, unimaginative setting, visually unappealing.
Settings:
Steps: 50,
Sampler: DDIM,
CFG scale: 6,
Seed: 123123123,
Model: Stoked-Illustration-v1-1_fp16-no-ema_BakedVAE,
Lora hashes: "add_detail_1.1: 895aa4d10ea0",
Version: v1.3.2
Aspect Ratio 1:1 - 512*512
Prompt 1
Prompt 2
Aspect Ratio 2:1 - 512*1024
Prompt 1
Prompt 2
Aspect Ratio 3:1 - 512*1536
Prompt 1
Prompt 2
Aspect Ratio 4:1 - 512*2048
Prompt 1
Prompt 2
Aspect Ratio 5:1 - 409*2045
Prompt 1
Prompt 2
Conclusion
Now obviously whatever I say will rightfully be seen as biased as it probably will be, as this is the model I find myself using the vast majority of the time lately. Even I however am surprised at how well it holds the composition together at the wider extremes. I also only see slight duplication in the image elements.
With this new info it looks as though my days off will be spent updating those awkward banners we all have on a site somewhere, the one that the sizing makes no sense whatsoever. Now I can just type in the size (or multiple/division of) and upload the result. Easy.
I hope someone out there gets something from this.
...lol after all that I have just realised the upload slots have nothing to do with the images. Every day a school day.