santa hat
deerdeer nosedeer glow
Sign In

Comfyroll

621
3.6k
30
Verified:
SafeTensor
Type
Checkpoint Merge
Stats
2,647
Reviews
Published
May 15, 2023
Base Model
SD 1.5
Hash
AutoV2
A82CD54E9E
default creator card background decoration
Holiday 2023: 5 lights
Akatsuzi's Avatar
Akatsuzi

The Comfyroll models were built for use with ComfyUI, but also produce good results on Auto1111. They currently comprises of a merge of 4 checkpoints.

The models can produce colorful high contrast images in a variety of illustration styles. A pseudo-HDR look can be easily produced using the template workflows provided for the models.

These checkpoint merges are particuarly good for:

  • semi-realistic styles (use Comyroll)

  • stylized anime (use ComyrollAnime)

  • illustration styles

  • 2D, 2.5D

  • NSFW

These merges were not designed for creating NSFW and in general do not produce unsafe images unless you deliberately prompt for them.

Comfyroll workflow templates:

A collection of workflow templates has been designed for use with these models. The collection includes workflows for anime and illustration styles.

https://civitai.com/models/59806/comfyroll-template-workflows

These templates were used to create the sample illustrations for the models. I have indicated in the comments on each image the template that was used and additional information such as the clip skip, noise masks and lora.

It is also planned to create example prompts for use in Auto1111.

Recommended settings (Comfy):

Sampler: dpmpp_2m karras, uni_pc normal

Steps: 10 steps for a contrasty manga-style look and 30+ steps for a smoother look

Clip Skip: both -1 and -2 both work well, -2 provides a good semi-realistic look, -1 is great for anime and illustration styles

Img2Img Denoise: usually between 0.85 to 0.95 (lower values are good for illustration styles)

VAE: vae-ft-mse-840000-ema-pruned (the models do not currently have baked VAE, so you will need a VAE loader node in your workflows)

These models work best when combined with noise masks in Img2Img or ControlNet. Please see the Comfyroll Template Workflows for examples of how to use noise. Recommended noise patterns include white noise, pink noise, dissolve noise and perlin noise.

Recommended settings (Auto 1111):

Hires fix is strongly recommended

Sampler: DPM++ 2M Karras, UniPC,

Steps: 20, Hires 30

Clip Skip: -2

VAE: vae-ft-mse-840000-ema-pruned

Prompting:

Here are some suggested keywords to use in prompts for these models:

  • Semi-Realistic Style

    • 2.5D, semi-realistic,

    • highest quality, detailed,

    • depth of field, natural blur, bokeh, HDR,

  • Anime Style (ComfyrollAnime)

    • 2D, 2.5D,

    • anime line-art, illustration,

  • Illustration Styles

    • 2D, 2.5D,

    • detailed concept drawing, illustration, line-art, stylized,

  • Lighting

    • subsurface scattering, chromatic lighting,

    • neon lighting, moody lighting, cinematic lighting,

  • Color

    • colorized, flat colors, limited color palette

  • Camera

    • from-above, from-below, from-side, from-behind,

    • portrait, cowboy-shot, close-up, long-shot,

Recommended LoRAs:

ukiyo-e 30% strength

NijiV5style with ComfyrollAnime, 50% strength

Longan/竜眼 Style with ComfyrollAnime, 70% strength

Resources for this model:

Textual inversions:

https://civitai.com/models/7808/easynegative

https://huggingface.co/yesyeahvh/bad-hands-5

Upscalers:

https://upscale.wiki/wiki/Model_Database

Links:

https://huggingface.co/Akatsuzi/Comfyroll

https://www.pixiv.net/en/users/17537676

https://www.deviantart.com/akatsuzi

https://twitter.com/Akatsuzi_AI