My first version of HyperNatural is here, I put a lot of effort and commitment into it, I will continue updating it and sharing it with you. I have refined the cues, hands, poses, and focused on realism to achieve what you can see in the sample images.
INDICATIONS:
-It is not necessary to use basic keywords like "HD, perfect, masterpiece, best quality, 8k, intricate, great detail, detailed face, detailed hands, detailed hair...etc, they are not noticeable changes in the final result.
-You can use simple negatives since it is quite well optimized. You can see the negatives used in my images to get an idea of the result.
-Add the ethnicity you are looking for in the prompt "American, Asian, Chinese, among other additions" if you are looking to create different faces.
-Lights, shadows and details are very good without requiring additional keywords.
-You can use "1girl" or "1boy", the model will vary the characters.
-The description of the face can vary or improve if you write features such as hair color, eyes, etc. in parentheses. It will turn out mostly good.
TIPS FOR USE:
-Steps: 20 to 30 (if the image has errors, please use higher steps)
-CFG scale: 5 to 8 (higher setting scale may lose realism)
-Sampler: 2M Karras Sampler, euler, euler a. Any Sampler (SDE, DPM-Sampler will result in more realism)
-Recommended size: 512x768 or 768x512, faster, then you can scale.
-Hire climber: ESRGAN_4x or LDSR (level 2x or 4x)
-VAE is not needed, if you use it you will get more vibrant colors.
Don't hesitate to leave reviews and samples of your results, I will be improving this model and sharing it with you, have fun creating!!
Description
V1.0 SafeTensor
FAQ
Comments (13)
You know how 'hyper' is spelled, right?... right? :)
Any info on how many images the model was trained on and how it does on wide shots?
The model will take about 200 images, and it will take about 8 to 10 hours to train. It defends itself perfectly in other more distant images, but civitai didn't let me upload them because it said they were sexy....if anyone knows how to do it, please explain it to me, it's the first time I've uploaded my model here, but I've seen others upload it Much more provocative photos, I hope it works well for you and you can upload an example. all the best.
@HyperNatural You can upload sexy images with no problem, just make sure you don't have a filter on in your settings. I've uploaded ton of sexy and even some full nude images without any issues.
As long as they're not depicting a celebrity or underage, you should be good to upload them.
It's really hard to judge a model from just face close-ups, as pretty much every model does well in that regard.
Have you compared your model to some of it's peers by generating the same seed and prompt with your model as well as other models?
@Jellon Yes, I have tried it and it works quite well for me, a good prompt is also important, I invite you to generate an image with this model and share it to see the result. thank you
This gives me almost identical results as the "The Truality Engine" checkpoint. Seems to be based on the same model. Both are equally bad at creating hands.
I do not know the model you are talking about and of course it is not based on it, it is true that the issue of the hands is complicated, I am working on "trying" to improve that, and I say try because I have not yet seen any model that does it perfect, if it is not a defect in the fingers, it is in the nails, it seems that the AI is not perfect yet. Thanks for your comment.
@HyperNatural Most models have a base model, I don't think you created something entirely from scratch. I did a model comparison and noticed that several of the images that came out were almost identical. The other model is a merge, yours is labeled as a trained model. So along the way I assume you used the same base for your checkpoint.
Fingers are very difficult to get right, even if you include several dedicated training images, it still comes out wrong in many cases. But that what ADetailer or inpainting is for.
What I really like with this model is that widescreen pictures come out very well. Better than in most other models.
@dbst17 *All checkpoints have a base model outside of ones like SD 1.5 or SDXL 1.0, etc. So no models are "trained" here other than those original 'base models'. It's all merges [fine-tuned models] even the so-called "trained" models from Civitai uploaders. Seeing overlap is to be expected since they will all share genetics, and possibly even more if said "trained" model used similar images in their dataset to fine-tune. It's not unlikely people would grab same images from same sources ;)
works great for me with my customs models and embeddings. thanks
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.