Pressing onto battle with the Valkyries

I'm not done with the valkyries yet. The prompt I am using:

winged angelic nordic viking valkyries descending from the sky, surrounded by a thunderstorm and lightning, wide angle, highly detailed cinematic headshot portrait, dramatic lighting, backlighting, atmospheric, beautiful women, painted by the pre-raphaelites, millais, waterhouse

The best seeds so far after thirty iterations:

3395687 Generated at 512x320 - which looks amazing.

So I'll take the seed and change the resolution, to see what the outcome / output ends up looking like. I want to see this in portrait format... Doubling the horizontal lines nets the following result:

That is using the same random seed, but with a very different result as the final outcome. I'm not sure I like it. Let's see what can be done to the following image by changing the aspect ratio.

3395699

How very different the other result is:

I'm now going to generate 100 portraits and choose the best five. This will be an exercise in absolute misery, because I hope I get more than five good results, which means I will have to be very particular, and very specific about the images I select.

It will be just like culling a Photoshoot. I remember those, in a time before I discovered AI art. If you want to see those, and what I used to do before I discovered this, go have a look at my collection.

My methodology here will be ordered. If I'm going to pick five out of 100 images, that means one from each batch of twenty.

The first twenty have now generated...

This doesn't fit the prompt at all. It is the most coherent of the images in the first twenty or so. I might change the rules. If the 2nd batch of twenty is terrible, or subsquent ones are, I can choose another image from the prior batches to sit in as one of the chosen ones, as the first few iterations of the 2nd lot of 20 don't yet look very promising.

7074645

7074678

7074690

7074671

7074734

There's two here that are really quite good, and those are the ones I'll be working on first. 7074678 has potential, but I'm concerned about the qualities in the face, I might need to do several bits of "fixing" there, depending on what the polished iteration ends up looking like after GFPGAN, and upscaling.

The face... still needs work, so I'll try using in painting again...

That isn't a great result. Time to try some alternative strategies, as I really like the baseline image here - the composition is strong, the pose of the character is strong, it is just let down by the unfortunately mangled facial expression.

Tackling some other in painting methodologies didn't work very well, but that's okay, because I can always generate another large bunch of images to see if I can find some better seeds, or perhaps, move the seeds around by a few digits to see how this influences the results.

Doing that did not yield any results that were remotely similar to the original output. That is upsetting, but I will... look onwards and search for more seeds that I can use.

This one isn't awful, but it is far from depicting a battle scene. I'm pretty tired at this point of my evening (as I normally am) - so I will just have to keep trying again and again tomorrow. Perhaps I'm going to stick to landscape next time and try to get a prompt that involves an epic battle scene, and what will appear to be countless figures involved. I want it to be very dramatic and dynamic in terms of its composition.