Meet DogePoint, the first model for sensitive young men™!
To answer the "wat checkpoint?" question on all of my showcases, these EPS and VPred bases are available so you can reproduce everything 1:1 or better, there are no other specialized models for anime, 3D or realism, these models should cover it all.
These models are calibrated around Euler A, CFG 5, Normal/SGM Uniform schedulers and resolutions around 1280 and 1536, you can go beyond that but the fracture risk, quality increase and rendering times are just not worth it. The showcased images are latent upscaled but you can generate straight on the final resolution without significant loss, which I recommend if you're not familiar with upscaling configurations and implications that would need an entire article to list part of it.
The main goal here is to run LoRA/LoCon with the best quality and the least interference possible, so I do recommend running mainly Illustrious and NoobAI trained models and avoid everything else, even finetunes labeled as IL that might mess up your blocks, give you blurry and washed up images.
It's noteworthy to mention that the alpha models aren't 100% finished, some improvements on sharpness, color temperature and small changes are still needed to clean and wrap it up. Image feedback is appreciated and will be taken into account for improvements. Any suggestions are welcome, let me know if there's anything else you need uploaded or written like the old "legacy base" and an upscaling guide.
Description
Fixed some CLIP and img2img interactions broken on conversion
Adjusted noise and color warmth
Sharper upscaling
Overall cleaner and more accurate
FAQ
Comments (5)
Your beta version is identical to your alpha version, as proven by their identical hash codes. Have you maybe uploaded the wrong file?
Not the same as you can verify on block analysis, but now it's gonna be a problem and I'll roll it back to a single model 🫠
@DogeOfVenice The output of both versions is identical and more importantly the hash is identical. If the hash is identical, the files are 100% identical, there is no way around that.
Is this a hybrid model? NAI implies one thing, but it's labeled as Illustrious.
I've only updated the IL models for now since there's not really demand for Noob right now, I won't port the LoRAs to Noob and they had a smaller texture range. Still you can run Noob models on it with a slight mismatch or seamlessly if you know what you're doing. They were what I had in mind last year (year this post was made) but these current versions make everything else pretty much redundant.