-
Sign up or login, and you'll have full access to opportunities of forum.
Discussion about A.I.
- Thread starter Doragon
- Start date
fallenmystic
Tribune
I think they look more realistic but still feel like enhanced Daz3D renders than photos, which is perfectly fine if that's the style you are looking for. But if you are aiming for semi-photorealism or feel annoyed at the flaws it introduces, I recommend switching to more flexible tools like Foocus, A1111, or ComfyUI (preferably with the Krita plugin, which is how I make my images).
If you don't have a graphics card with 10GB+ VRAM, there are several options for running such tools on a GPU cloud. Online services like Krea are quite limited in the features they provide, so they are often insufficient for serious work.
With the tools I mentioned above, you can create a series and fix the flaws. The former is a bit tricky, but it's definitely doable with the right tool (e.g. IPAdapter / Face ID) and workflow.
As for depicting real people in AI images, changing the backdrop of a photo or making photo manip work realistic is quite trivial to achieve. You can use the same method I just mentioned for something more exciting too. To make a series, you'll need a way to preserve the identity of characters across the images. And if it's possible, it means you can preserve the face of you and your friends across the original photographs and AI-generated images also.
fallenmystic
Tribune
I heard that a new version of the Krita AI plugin had just been released, which added a regional prompting feature, so I tried it out in conjunction with the Pony model I mentioned above.
To make a long story short, I think I've found the ultimate AI tool I've been looking for:
To share my new workflow:
I'll link a video to help you get an idea of how this feature works:
To make a long story short, I think I've found the ultimate AI tool I've been looking for:
To share my new workflow:
- Live regional painting with Pony Lightning to visualise my idea.
- Inpaint to fix deformations and add details to create a base sketch.
- Make controlnet images out of 2.
- Switch to the generation mode (i.e. non-live mode) and do "seed hunting".
- Adjust things like CFG, sample counts, and so on with the seed found in the previous step.
- Inpaint any remaining problems, if necessary.
- Upscale the result using SUPIR.
- Post-process in a photo editor (e.g. Darktable) for finishing touches (e.g. colour correction, adding grains, etc.).
I'll link a video to help you get an idea of how this feature works:
fallenmystic
Tribune
As my honeymoon period with the Pony Diffusion was over, I began to see both its strengths and weaknesses.
I'm still very happy with its ability to depict various lewd concepts with relative precision. It really helps when the model understands tags like "grabbing_another's_breast," for example. I even stopped using my own slave restraints Lora, because the Pony model does it better by default.
On the flip side, I often found its anime origin sips through, even though I'm using a photorealistic variant. This was a typical image I got while I was seed-hunting for my latest render:
As you can see, things like the face or the suit clearly show strong anime/illustration influences. I envy those who create anime-style works since the Pony Diffusion would be a gift from god to them as it is (I have to wonder how gods would think about censorship, though). But as I'm aiming for photorealism, I must find a better way of doing it.
I decided to run it through a non-Pony model, improving realism while destroying the details. The lost details could be recovered by inpainting with the Pony model again since it's much better at keeping with the overall image style when used that way.
There's something which might be worth mentioning that I tried for the first time in my latest render. The original base image looked like this, which I felt rather dull:
I wanted to make it look like they were on a stage with spotlights on the girl's body. It'd make the scene more dramatic while emphasising the girl's humiliation. The problem is that precisely controlling the lighting condition might be very difficult in Stable Diffusion. Well, at least it was so until recently.
Now, we have IC-Light, which allows us to change the lighting condition of an image without destroying its content. With IC-Light, I could add a nice spotlight on the girl's body like this:
On a side note, similar AI technology seems to be making a large impact on filmmaking as well since it allows filmmakers to freely experiment with different lighting setups in the editing room after the shot is taken.
Anyway, I'm very happy with my current Pony / Krita + SUPIR/Light-IC/etc workflow. I used to spend a few days per image before, but now I can generate better ones just in several hours. I still don't know what will become of SD3, but it looks like I can still squeeze a bit more fun out of SDXL while I'm waiting for the next big thing.
I'm still very happy with its ability to depict various lewd concepts with relative precision. It really helps when the model understands tags like "grabbing_another's_breast," for example. I even stopped using my own slave restraints Lora, because the Pony model does it better by default.
On the flip side, I often found its anime origin sips through, even though I'm using a photorealistic variant. This was a typical image I got while I was seed-hunting for my latest render:
As you can see, things like the face or the suit clearly show strong anime/illustration influences. I envy those who create anime-style works since the Pony Diffusion would be a gift from god to them as it is (I have to wonder how gods would think about censorship, though). But as I'm aiming for photorealism, I must find a better way of doing it.
I decided to run it through a non-Pony model, improving realism while destroying the details. The lost details could be recovered by inpainting with the Pony model again since it's much better at keeping with the overall image style when used that way.
There's something which might be worth mentioning that I tried for the first time in my latest render. The original base image looked like this, which I felt rather dull:
I wanted to make it look like they were on a stage with spotlights on the girl's body. It'd make the scene more dramatic while emphasising the girl's humiliation. The problem is that precisely controlling the lighting condition might be very difficult in Stable Diffusion. Well, at least it was so until recently.
Now, we have IC-Light, which allows us to change the lighting condition of an image without destroying its content. With IC-Light, I could add a nice spotlight on the girl's body like this:
On a side note, similar AI technology seems to be making a large impact on filmmaking as well since it allows filmmakers to freely experiment with different lighting setups in the editing room after the shot is taken.
Anyway, I'm very happy with my current Pony / Krita + SUPIR/Light-IC/etc workflow. I used to spend a few days per image before, but now I can generate better ones just in several hours. I still don't know what will become of SD3, but it looks like I can still squeeze a bit more fun out of SDXL while I'm waiting for the next big thing.
fallenmystic
Tribune
Amid the chaos and disappointment after SD3's disastrous release, a glimpse of hope finally emerged:
Civita's announcement: https://civitai.com/articles/5884
InvokeAI's announcement: https://www.invoke.com/the-open-model-initiative
LAION and Comfy are also on board but haven't yet made an announcement.
I've already mentioned many times in this thread why upholding the open-source ideal is essential for AI development, so I won't repeat it here.
So, to summarise the news, a joint effort to develop an open-source AI model for generating images, videos, and audio was just launched.
While several other attempts with a similar goal have been made before (including "Unstable Diffusion," which turns out to be more or less a scam), The Open Model Initiative may be more successful than its predecessors because it was led by those who have the actual knowledge and resources to create capable community models. With Civitai, Comfy, InvokeAI, and LAION all participating in the project, they already have skills, training data, and even the community to support such an endeavour.
It'll probably take quite some time before we can see anything concrete coming out of the project.
But I'm cautiously optimistic about the prospect, considering the members participating in the initiative.
The Open Model Initiative
Civita's announcement: https://civitai.com/articles/5884
InvokeAI's announcement: https://www.invoke.com/the-open-model-initiative
LAION and Comfy are also on board but haven't yet made an announcement.
I've already mentioned many times in this thread why upholding the open-source ideal is essential for AI development, so I won't repeat it here.
So, to summarise the news, a joint effort to develop an open-source AI model for generating images, videos, and audio was just launched.
While several other attempts with a similar goal have been made before (including "Unstable Diffusion," which turns out to be more or less a scam), The Open Model Initiative may be more successful than its predecessors because it was led by those who have the actual knowledge and resources to create capable community models. With Civitai, Comfy, InvokeAI, and LAION all participating in the project, they already have skills, training data, and even the community to support such an endeavour.
It'll probably take quite some time before we can see anything concrete coming out of the project.
But I'm cautiously optimistic about the prospect, considering the members participating in the initiative.