- Joined
- Oct 20, 2019
So it's not convenient for me to fire up runpod right now but I can offer some suggestions.
The first is one you probably don't want to here but if that's the level of control you require, you need to take the next step in your learning and try out either image to image or control net. You'll find image to image the simplest as you are just taking blank latent image and using an actual image instead and then experimenting with the steps and denoising to find a good balance between blitzing everything and changing nothing. I know that ComfyUI looks more complex but I actually find it simpler because you can see and understand the steps better. There's an example work flow in their examples folder which is actually just a PNG you drag onto the UI and it will replicate the image to image workflow. Image to image can let you tweak things or modify an image so if you had say drawn what you wanted you could use AI to make it more impressive. There's also in-painting which can change small parts of the pic such as adding a blindfold. That's something of an art but it means you could create your image in stages, adding elements as you go. But it's tricky and honestly what you probably want for something like this is Control Net (which isn't available for Flux yet but they say is imminent). Controlnet will let you fine-control poses (amongst other things):
But enough of the accurate and tough answers. There are things that you want to hear as well which is just things you can change about your prompt.
Night time moody lighting, a woman in all black, eyes blindfolded with a black scarf, using both hands holding an ancient broadsword pointing up, with both hands on the hilt holding it up, surrounded by darkness, there is no forest, there is no beach, she is skipping atop water in a pond with waves dispersing at each graceful ballerina step from left to right, the water is clear and beautiful, atmospheric dark shot, camera is looking down at her.
Instead of "a woman in all black" be specific: "woman in a black dress" If you need to add things like "black ballerina shoes" then add them and re-run as you go. Try to keep prompts simple at first and maybe add detail later. So instead of "using both hands holding an ancient broadsword pointing up, with both hands on the hilt holding it up" with its multiple holdings and "using", just try "holding a sword above her head with both hands"
Part of your problem is slightly conflicting requirements. Example: blank background, standing on a clear and beautiful pond, camera looking down at her, camera... Lets put these together. If you're looking down at her then the background must show the water. But you've also told it to have a blank background. So I am inferring that you want something that is more like an illustration - like a drawing of all this on an otherwise white canvas. I don't know if I'm right about what you want but it you did want a photograph then "blank background" is going to be a little odd for something that includes nature elements. So if you want more of a drawing, lets add "drawing of..." at the start here. And we should also consider what you mean by "night time". Ordinarily that might mean all sorts of things - if the background is an alley then it might mean street lighting, if it's rural it might mean stars and moon in the sky. But with "blank background" it can mean none of that. So you're trying to set the model against itself to a degree. Same with clear water. Water isn't clear at night - it's either a black void to look upon or it's a mess of reflections. But it's not something you look down through. So what exactly are you expecting. If it goes with "clear and beautiful" water which it probably will, it'll give you something that looks like daytime in the water, at least. So to help you, I would suggest you think about what you want specifically and then request that directly so there's less room for interpretation. "dark green water", "reflections of stars", whatever.
I'm fairly certain this image isn't what you want but it might provide you a starting point that you can begin adding things to. Note I went with the idea that you wanted a drawing due to the blank background requirement and that you didn't specify anything like "highly detailed, highly realistic, photograph" in your prompt.
So this prompt got me the image below: "drawing of a woman in black dress wearing black blindfold holding a sword above her head with both hands. She is walking on water, blank background, viewed from above, sword pointing upwards, night, atmospheric"
For comparison I did the same prompt as a photo so you could see what I mean about "moody" and "nighttime".
(Prompt was: "photograph of a woman in black dress wearing black blindfold holding a sword above her head with both hands. She is walking on water, blank background, viewed from above, sword pointing upwards, night, atmospheric")
As you can see, it's not really sure how to make it nighttime or atmospheric with nothing to work with to do so. You will have to explain how you want it to be night or moody to help it out if you want a blank background. Like if you want a dark background, tell it that.
Unfortunately you're probably just using a web API in which case it may not support re-using the same seed. Depending on your programming knowledge, I would look at using the API programatically. Failing that, run it locally. (Unless I've misunderstood and you already are). In which case you can used a Fixed seed and see the impact of your prompt changes only without other random changes.
So I hope that was on some help. The real answer is ControlNet. But there's still some changes you can make to your prompt to increase your chances.
Good Lord - the AI is an admirer of Robbie Burns!
