Hello!
I've uploaded 12 models today and those have different samples than usual. I got bored of the extreme close-ups and tried to invigorate the prompts with new photographer styles as well as new locations.
On top of that, I've used ADetailer - a real gem when it comes to SD generation!
It's an A1111 extension (not sure about ComfyUI) available at: https://github.com/Bing-su/adetailer
It is very easy to use, there is no configuration needed (pretty much) and the results are magnificent.
Some people asked me - "your models are great at close up but when I try to have a full body shot the quality goes down, what can I do?"
My answer was usually: do a high.res.fix with a higher denoise level (starting from 0.2 for close-ups, through 0.4 for half-body shots up to 0.6 for full-body shots).
Now, we have a new player in town (well, we had it for some time but not everyone used it, me included). The default value of inpainting 0.4 is good enough. You can take a look at my recent models to see them in action.
When people asked me: hey, your models are great at close-ups, but the far-away shots are bad. My response was: this is the way SD works. You could train a model on some full-body shots but it will still be lacking. So I decided to focus on making great faces and full body shots were delegated to inpainting at first, and later to ControlNet (which is still great!). But now you can use ADetailer in addition to all that and it will pretty much serve as automatic inpainting. I was really positively surprised with the result - if you haven't played with it yet - I strongly encourage you to do so!
Roop: I have finally updated my A1111 and installed Roop. So far I've only played with images (no movies/animations) and I have mixed feelings about it. It is kinda cool, you can get some likeness for a person without having their model.
But since I am into training models, I've checked how models + roop & original photos work.
And I have to say they work pretty well. The flow is as follows:
You make a normal txt2img output WITHOUT hi.res.fix (because the only available onxx model operates on 128x128 resolution; there are rumors about 512x512 model but it has never been shared publicly, which is a real shame!).
So, you have Roop enabled (and you can have the upscaler enabled). You generate the image and a normal txt2img process happens but then Roop works its magic and replaces the original person with the one from your image. (there is also a way to use roop for injecting more than one face in the image, take a look at this video from Sebastian Kamph for more details:
)
Going back to single-person generation, if you have any model of that person (Dreambooth/LoRA/LyCORIS) I suggest you move that output image to img2img and rerun it with 0.2 - 0.3 denoise level. This will improve the overall quality and the likeness of the person (but you have to use a really got image of the original person to begin with!)
At that denoise level, your model will handle nicely the low-resolution artifacts and won't change that much the face that Roop made.
I haven't yet tried the animation, which is where the roop really shines but when I do - I will definitely let you know!
As for the content I produce, I would like to know what are YOUR expectations. What would you l like to see the most?
Here is the survey (multiple choice): https://q65fh4mdq.supersurvey.com/#
Would be great if you could just click on the stuff you would like to see (it's anonymous, just click and forget :P).
Shameless plug, as usual, if you feel like I am contributing to the community, you can support me by donating over https://www.buymeacoffee.com/malcolmrey
I keep making 1.5 and SDXL models.
P.S. A user suggested the following negative embeddings: BadDream, (UnrealisticDream:1.2), realisticvision-negative-embedding
I am running with those pretty often and they seem to produce higher-quality images (and fewer images that have some weird artifacts).
However, I have noticed that those 3 embeddings have a negative effect on the likeness of specific subjects.
So it is good to check if the person's token produces good results before adding those negative tokens.
I have an idea for another article (for 2-3 weeks already) about a way to produce a better likeness of a subject. I've been testing this for over two weeks now and I am confident that it works. My plan is to produce an article about this way this week so keep tabs on it :)
Cheers!