Qwen edit can take an image as a sample, and work with that. “The character in a victorious pose” would get whatever character you have, and reproduce it in a victorious pose. Couple of examples:

And a little janky because it IS generative AI after all…

Edit: and a bonus screenshot showing how little effort I had to put into this lol

That sounds so hard to do locally. I don't even know where I would begin. Is that expert level stable diffusion?

As far as I know, one way to do it is to use an image2image model / image editing model. You'd need to generate one reference image of your character. And then feed the reference image into it and tell it to draw that character in a different scenario. I haven't done it in a while, not sure what people use these days to do it locally.

My limited experience is that stable characters across a number of images are a weakness today, and I wouldn't be confident that genAI is a great way to go about it. If you want to try it, here's what I'd go with:

• If you can get images with consistent outlines via some other route, you can try using ControlNet to do the rest of the image.

• If you just need slight variations on a particular image, you can use inpainting to regenerate the relevant portions (e.g. an image with a series of different expressions).

• If you want to work from a prompt, try picking a real-life person or character as a starting point, that may help, as models have been trained on them. Best is if you can get them at once point in time (e.g. "actor in popularmovie"). If you have a character description that you're slapping into each prompt, only describe elements that are actually visible in a given image.

I've found that a consistent theme is something that is much more achievable, in that you can add "by " to your prompt terms for any artist that the model has been trained on a number of images from. If you're using a model that supports prompt term weighting (e.g. Stable Diffusion), you can increase the weight here to increase the strength of the effect. Flux doesn't support prompt term weighting (though it's really aimed at photographic images anyway). It's possible to blend multiple artists or genres as prompt terms.

Damn, probably won't run on a 3090 anymore then, Flux 1 barely fit on it

It's significantly larger than FLUX.1 (32b/64.4 GB vs 12b/23.8 GB) so I would hope it's more than a small iteration.

Still that terrible license, won't let us provide it on the horde

Huh, seems interesting, no big claims or anything on the page. Looks like a small iteration

Cool! I'm gonna try it out. Thanks!

Whoops. Thanks for that.

The blog link has an extra dash on the end (or it did when I posted this comment, at least)

Fixed link: https://rocm.blogs.amd.com/artificial-intelligence/nitro-e/README.html

I noticed that too. The one with the woman, her head stays pretty close to the same left facing profile shot from the reference, but at least her body changes.

Honestly, this refinement seems like it's gaming the metric a bit. I have no doubt there's innovation here, don't get me wrong, but the first two examples are not just retaining the character, they seem like overfitting.

The model in many of the images reproduces the exact same armor angle, shield angle, head/helmet angle, etc, without any modification - i.e., it isn't learning how to reproduce a persistent character in multiple angles, scenes, etc, just the exact illustration fed into it, while changing the setting and positioning.

I think that there's maybe a need for something like this, but if it's not just going to be a one-off research project

which maybe this is, which is okay

I'd very visibly version the testset and its results from the get-go. You're going to want to add more tests to it over time, and it'll affect change test results, and you're going to want to be able to reproduce results.

That's pretty impressive that they show it from the 2nd grader blob sketches to the finished animation.

This really does look like a signal to where the future workflows will be, because careful key-framing addresses a lot of artificiality inherent in AI video.

Thank you

https://dontasktoask.com/

!imageai@sh.itjust.works

Just use Google Gemini

No dog no edit.

Not saying I will do it, but at least we will have something to look at.

Thanks. I forgot I had Bypass Paywalls Clean installed.

Without paywall: https://archive.is/4oEi2

Yes but not everyone has something that can run a local model. We also specify exact versions of the local models we run so people can replicate locally if they want. Still don't see why your approach dismisses the whole extension.

Any reliance on remote compute is fragile, because you lack control of the model. Exact versions matter. Every thread that goes 'how do I use the old [service name]?' is someone learning this lesson, often too late.

What's your problem with the ai horde?