Search Results
8/1/2025, 9:59:01 AM
>>102965610
When you mention forest without 'sky, clouds' you are limiting your field of view which will always get cut off by trees in the background. Try adding those tags I mentioned too. They might help. 'photo background, real world location'
I tried your model and prompt on comfy but I had to turn off your weight numbers because comfy does not take well to any numbers above 1.3 like (censored:3) and turn off clip skip -1 and these were my results:
No lora:
https://files.catbox.moe/pdbql1.png
Lora at 0.3:
https://files.catbox.moe/y7unql.png
Mileage may vary but as another anon mentioned before, negatives can really drag down images. Are you getting chibis without prompting for them? If not, then that can be removed. Monochrome/grayscale images without mentioning 'monochrome' in positives? Removed. Dark-skinned male at 4? There is no male. Removed. Webui should allow you to save your current negatives in some form or another and you can get them back with a single click later when you need them for nsfw. I believe this was like a dropdown or something, under the generate button.
>>102971817
Funny part is this was extracted from a failed merge with the 2dn model while I was trying to get a model that can do Realistic backgrounds and Anime faces (hence RBAF) mainly because the resulting model started outputting pure black and pure white 0,0,0-255,255,255 without being a vpred model. As you're already on a vpred model I am not sure what benefits it can bring. It could simply be a case of "works on my machine". I would try it under <0.5 strength and with a lot of background mashing.
> architecture, building, castle, cloud, cloudy sky, dilapidated, fantasy, fog, foliage, glowing, grass, lamppost, , outdoors, overgrown, reflection, ruins, , sky, sunlight, tower
You can search on danbooru for 'no_humans scenery' for ideas but avoid using those two tags in your prompt if possible.
When you mention forest without 'sky, clouds' you are limiting your field of view which will always get cut off by trees in the background. Try adding those tags I mentioned too. They might help. 'photo background, real world location'
I tried your model and prompt on comfy but I had to turn off your weight numbers because comfy does not take well to any numbers above 1.3 like (censored:3) and turn off clip skip -1 and these were my results:
No lora:
https://files.catbox.moe/pdbql1.png
Lora at 0.3:
https://files.catbox.moe/y7unql.png
Mileage may vary but as another anon mentioned before, negatives can really drag down images. Are you getting chibis without prompting for them? If not, then that can be removed. Monochrome/grayscale images without mentioning 'monochrome' in positives? Removed. Dark-skinned male at 4? There is no male. Removed. Webui should allow you to save your current negatives in some form or another and you can get them back with a single click later when you need them for nsfw. I believe this was like a dropdown or something, under the generate button.
>>102971817
Funny part is this was extracted from a failed merge with the 2dn model while I was trying to get a model that can do Realistic backgrounds and Anime faces (hence RBAF) mainly because the resulting model started outputting pure black and pure white 0,0,0-255,255,255 without being a vpred model. As you're already on a vpred model I am not sure what benefits it can bring. It could simply be a case of "works on my machine". I would try it under <0.5 strength and with a lot of background mashing.
> architecture, building, castle, cloud, cloudy sky, dilapidated, fantasy, fog, foliage, glowing, grass, lamppost, , outdoors, overgrown, reflection, ruins, , sky, sunlight, tower
You can search on danbooru for 'no_humans scenery' for ideas but avoid using those two tags in your prompt if possible.
Page 1