I trained 10 LoRa's on 1300 realistic photos of very high quality at 1024x1024 resolution and mixed them with the Holygenex model, then combined the resulting model with the SD 1.5 base model.

The resulting Holygenex LR model can generate 768x768, 1024x768, 768x1024, even 1024x1024.

The task is to create a model which can generate high resolution images with excellent quality on the SD 1.5 engine.

From now on, you can say that SD 1.5 can generate images with 768x768 and 1024x1024 resolution.

To get a good quality image you need to use Hires. fix. There is no other way!

The images in the examples are generated at 90-150 steps and Hires. Fix with Denoising strength: 0.3 using Upscale by 2 with best upscaler 4x-UniScale-Balanced [72000g]. See the full picture descriptions in the examples. Click "Copy generation data", copy the text to notepad and see the full description.

• For panoramic images it is better to use a resolution of 1024x768

• For portraits you can try 768x960 or 768x1024

• For precise geometric objects 768x768

If you have graphics cards with 8GB memory or more, use Hires. fix. The quality and detail of the image improves several times over.

If you have weak graphics cards, you can generate portraits without Hires. fix, use a resolution of 768x960 or 1024x1024 and check the Restore faces.

For best results use the VAE, vae-ft-mse-840000-ema-pruned https://huggingface.co/stabilityai/sd-vae-ft-mse-original/resolve/main/vae-ft-mse-840000-ema-pruned.ckpt

This embeddings must be used in your NEGATIVE prompt:

bad-hands-5 https://huggingface.co/nolanaatama/embeddings/blob/main/bad-hands-5.pt

EasyNegative https://civarchive.com/models/7808/easynegative

Negative Prompt for example:

EasyNegative, bad-hands-5, (((nude, naked,child, child face))), (out of frame:1.1), worst quality, low quality, jpeg artifacts, cgi, 3d, render, sketch, cartoon, drawing, ugly eyes, (out of frame:1.1), worst quality, low quality