Sign In

Fruit Explosion

423
1.7k
23
Verified:
SafeTensor
Type
Checkpoint Merge
Stats
1,735
Reviews
Published
Jun 8, 2023
Base Model
SD 1.5
Hash
AutoV2
3CA529FB14

This is intended for generating detailed anime style images in wide, tall, or square images. It has been tested with the following ratios: 16:9, 9:16, 3:2, 2:3, 4:3, 3:4

Note that things can get weird when the initial resolution is fairly high without using high res fix. Since that adds a significant amount of time for those I typically start with a max of 768 along one side, though that can get wonky sometimes due to how Stable Diffusion works.

For example, I like my 2:3 images to be 512x768. After doing whatever initial inpainting I find necessary, I use Ultimate SD Upscaler to increase resolution. It may take some testing with the denoising strength, etc.

Experiment with upscalers to see which gives a better result for the project in particular.

Upscalers I like:

Foolhardy Remacri

NMKD Siax 200k

NMKD Superscale SP 178000 G

UltraSharp

lollypop

4x-UltraMix Restore

Please do not hesitate to ask questions. I am by no means an expert, but I will try to help if I can.

Note that the images above were upscaled in img2img a couple times using Ultimate SD Upscaler (see link above) using ControlNet tiling. I do tweak the prompts depending on the image, but as a general test the first img2img upscale removes the negative prompt without changing the positive prompt and upscales x2 with denoise strength around 0.2-0.4. After tweaking it from there, that image is upscaled again x2 with no prompts at all at about 0.05-0.1 denoise strength using ControlNet tiling once again.

ALSO note that some images do not have the sampler listed since they used DPM++ 2M SDE Karras and that is not on the list at civitai yet.