Time to summon some new valkyries to add to my collection of images in this theme. I am using the Stable Diffusion Web UI to generate these images, and even though I requested random seeds, it appears to be generating the seeds in a sequential order.
I have requested fifty images with the following prompt and other settings.
The prompt:
winged Valkyries 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, water house
- Resolution 1024x640
- 40 inference steps
- 7.5 guidance scale
- GFPGAN ON
Shortlisting 1 in 10 generated images
These are the short list. I am biasing towards "sfw" images here, as the prompt somehow managed to get some bare torsos. The representations were artistic, but not as of high quality as the ones I've selected here.
5338112
Interesting composition, and the detail in the background remains coherent. I am concerned about the quality and resolution of the detail in the faces of each of the valkyries here, but this is a good candidate!
5338113
This is a very dynamic composition, and I like the way the figure on the right hand side of the frame appears to be levitating. There's an ethereal and mysterious aura with this one, it may look better with further coherence, but I also love the atmosphere generated by the clouds.
5338120
This looks like the conclusion of a battle. This is probably my favourite one so far. I like the Valkyrie almost getting lost in the clouds to to the top side of the frame, and the way that the character's attention is following that direction.
5338125
This is unqiue. One character is embracing the other, though the arm featured in the embrace appears to be a little bit disembodied. The other character beckoning to someone out of the frame is also another point of tension to add to the image. The boording weather and the swamp / lake in the background is a nice detail.
5338149
I short listed this one on account of the interesting facial expressions.
From the above images, the seed I like the most is 5338120, so this is going to get a higher number of inference steps to see what the outcome of a more resolved or coherent image may be.
I doubled the inference steps to 80, and I think the result is worse.
As a result, I will drop to 50, as some middle point between 40 and 80.. but I'll do it slowly. The initial seed and inference numbers got a brilliant outcome.
Completely different again - so I'll drop to 45, to see if further coherence can be gained. And... it is a no. It's the same, fundamentally as the above image, a complete departure from the image that is generated with 40 iterations.
Increasing beyond this generates an image similar to the 80 one, so 40 must be the sweet spot. That was using the euluer_a sampler, I'm going to see what happens with PLMS.
You'll notice the following images are all pretty much the same, with small variants, owing to the different samplers being used. Feel free to treat this as a reference piece.
ddim sampler
plms sampler
heun sampler
euler sampler
euler ancestral sampler
A much more artistic result, back to the original image that I loved. What will the remaining samplers result with?
dpm2 sampler
dpm2 ancestral sampler
lms sampler
This one was very quick in terms of generation.
We can see that every sampler except euler ancestral is pretty similar, and the result from euler ancestral is the best, most artistic one for this particular vision. I am going to try (again) to up the inference count on that one to see if I can get that particular scene to become more resolved. Otherwise, I will be employing some other strategies to try and get some more coherence and veracity into the image.
euler ancestral sampler with 200 inference steps
What a different result, yet again. It appears that the magic mix is that 40 steps to get that specific image. This is very interesting to me. Given the output, I'll take the initial image that I liked into another algorithm to try and resolve out more detail.
Using the image (and the same prompt) as input, I want to see what the output becomes with a new, single random seed. Same sort of composition, but again, an entirely different result:
Over now to a different algorithm to see what I can get out of some alternate sampling and in painting options. This is going to have a slightly lower guidance scale - so may result in a different result.
Hmm. The results using Codeformer for fixing faces are ... different.
I'm not sure I like the output, reversing the direction of the gaze of the character to the left. Perhaps it is better being "wish-washy" with limited coherence, allowing my eye to fill the gap. Which of the seeds were your favourite?
I hope the examination of the different samplers is useful for someone out there. I had fun exploring this seed, and the different samplers, and am likely to continue on this track until some other idea derails me.
Want more content from me?
Witness my futile efforts to play my Steam Game collection in alphabetical order.
Are you aware that I love photography? Check out my work in a collection.
If you want to see my Splinterlands antics and rants live, Find me on Twitch
If you prefer sleeping in your designated time zone, go watch replays on YouTube.