@VelvetS So basically GPT2 images has extremely bad watermarking that it doesn't hide sufficently and therefore is visible in the final image. You can see it as strange dotted/noise/grid patterns throughout the image (especially in backgrounds, clouds, fire etc.) Here are examples where its obvious what I mean (although it is not THIS bad anymore)
- https://www.reddit.com/r/ChatGPT/comments/1stvpeq/chatgpt_image_2_is_broken_weird_artifacts_90_of/
- https://www.reddit.com/r/ChatGPT/comments/1sv98ga/gpt_image_2_keeps_adding_weird_tiling/

What I meant is I can see this pattern in some of the images generated with this lora. For example all images where there is fire you can clearly see pattern in the fire. Worst example is the image with the tiefling where you can see dots through the entire image (masked for some parts of the image as embers, in other parts as black dots on the shirt of her arm, the strange metal spots on her shoulder armor that really dont make sense or the grid pattern of the fire.

In the image with the snow wolf you can see the pattern it tries to hide throughout the entire image as well (mane of the wolf, snow particles, weird dots on the armor, etc.)

Or as a final exmaple in the image with the snow queen you can see the pattern and dots it tried to "hide" in the embroidery of her dress or the "ice" of her weapon.

I would assume its hard to fix since its a hard limitation of GPT2 images itself

@VelvetS Strange. But its all good. Its managable anyways as it doesn't appear in every image and isn't nearly as noticable. And the style is worth a little bit of fiddeling around 👍