I decided to make a more realistic model based on my other model, but I confess that I don't have much experience with photorealistic images, so I tried to replicate the best possible way of a model similar to the realistic one, but still with 3D elements, to make it easier. in the use of prompts. This model is in the testing phase and I don't know if it is creating images as good as my 3D model.
To merge this model, I used the models below.
DarkPhoenix 3d Real: https://civarchive.com/models/172393
Realspice: https://civarchive.com/models/158734?modelVersionId=208629
Analog Madness: https://civarchive.com/models/8030?modelVersionId=160495
AbsoluteReality: https://civarchive.com/models/81458?modelVersionId=132760
For this model I recommend using some basic prompts like positive and negative.
Positive prompts: masterpiece, high quality, best quality, Detailed, Intricate, Realistic
Negative prompts: EasyNegativeV2, low quality, normal quality, worst quality, (bad anatomy), painting, illustration
Other combinations of negative prompts also work, as does using Fastnegative or BadDream (but I strongly recommend using EasynegativeV2, because of all my tests, this is the negative that works best without damaging the image too much in StableDiffusion)
Don't use Embedding NegfeetV2, I don't know why, but it ruins the image if you use this Embedding in this model
Recommended settings
Sampling steps: 20-30 (Only use more steps as a last resort for very complex images, as an excessive number of steps tends to ruin the image)
CFG Scale: 6-10
Hires. Fix
Upscaler: 4xNMKDSuperscale_4xNMKDSuperscale
Hires Steps: 15-20
Denoising Strenght: 0,20-0,40 (The more force, the cleaner the image, but the more the original quality is lost.)
Upscale By: 1,5 - 2,0 (512x512: 2.0, 512x768: x1.8, 512x912: x1.5)
Adetailer:
Adetailer Model: face_yolov8n.pt
Inpaint denoising strength: 0,25 - 0,40
ADetailer steps: 15 - 20
Did you like the content? I would be grateful to receive a coffee
https://ko-fi.com/berserkfg
This model is open source for pornography, therefore pornographic images are at the image creators' own risk, and the model will not be held responsible for any image with heinous content.
Description
Extremely early version for testing, so I'm aware of the bugs the model may have. I will be happy to receive opinions and criticisms about the model.
FAQ
Comments (14)
Trying to figure out what I'm doing wrong here. I tried using the same prompts and metadata from one of your images and got a weird doll thing. I've tried prompting images and all I get is a smoky image or something seriously messed up with a lot of artifacts. I'm using all the same settings and have tried prompting in the same style, however, this model doesn't seem to do as intended?
This is strange, it should be possible to replicate it, have you tried using one of my images and playing it in PNG Info? You will already have all the information I use in Stable Diffusion, and from there you can send it to txt2img and replicate the image
Now if you are using Civitai to replicate the image, I can't say how it is working here, as I have never tested my models here in Civitai to see how they are working
another thing, all my images I use Hires Fix and Adetailer
This information will appear in PNG Info, if you place the image there
I'm only getting NaNs or black images. I think something is seriously wrong with the model.
This behavior of the model is quite strange, but not disposable, as it was version 0.1 and I knew that bugs could appear (a problem that no one had reported this to me yet). I believe version 1.0 will resolve this.
This behavior of the model is quite strange, but not disposable, as it was version 0.1 and I knew that bugs could appear (a problem that no one had reported this to me yet). I believe version 1.0 will resolve this. I have been carrying out tests to update this model to version 1.0, a problem that I have not yet been able to find any improvement in the images (every test I do, it seems that the images are inferior to the current version)
I'll try to update it this week.
@BerserkFG the checkpoint's Unet and VAE shows up as broken, it's also missing the text encoder completely. This is probably why you could find any improvements to your test images.
@Pitpe11 Yes, when I merged this model, I didn't test it much and after it was launched, I realized that the seeds were never possible to recreate the same images. But when launching version 1.2 of my other model, I discovered what I had done wrong, as version 1.1 of my other model had similar problems.
Cool, well the test images look amazing. I have been using the Darkphoenix 3d model and was/am highly interested in this one! looking forward to an update. :)
The initial idea for this model was to create a closer to realistic version of my Darkphoenix 3D model, but still be easy to create good images (using simple prompts), as I couldn't create very good fantasy images in the other realistic models
I believe I have good news about version 1.0, I did the tests and didn't find the bugs that were in version 0.1, and I believe I'm creating the cover images to post by 12/17
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.