Photo2Anime | ComfyUI workflow版本v1.0 (ID: 318079) 综合资源合集综合资源合集

(Note: in ComfyUI you can disable nodes with Ctrl+M. You can quickly disable whole groups with right click option. Useful to turn off parts you don't want.)

This workflow takes a realistic image to generate 5 different anime versions of it in. Works with real photos!

Reverse workflow: Anime2Photo

Comparison of results

→ full size image here ←

Installation and dependencies

Install WAS Node Suite custom nodes;
(optional) Install WD 1.4 Tagger custom node;
(optional) Install SD Prompt Reader custom node;
Download and open this workflow;
(kind of important) Check "Resources and substitutions" section below for recommended extras.

You only need WD 1.4 Tagger if you want to interrogate image for prompt automatically.

You only need SD Prompt Reader to get prompts from metadata of realistic generations.

With neither, you'd have to write the prompt for picture manually.

Resources and substitutions

Workflow was made with possibility to tune with your favorite models in mind. There is the node called "Quality prefix" near every model loader. Just put most suitable universal keywords for the model in positive (1st string) and negative (2nd string). Note that KSampler parameters are thoroughly tested and fine-tuned for selected models.

Semi-realistic model (intermediate): XenoGASM. How to substitite: with semi-realistic model of your choise, like CT, EighthThing or Lyriel. Add anime-guiding keywords to "Quality prefix" node.

3D Anime model: OnlyAnime. How to substitite: with anime model of your style choise. Examples of 3D anime models: Kotosmix, xRikishi's Unstable Diffusion, AniVerse. Add universal quality tags recommended by the model author to "Quality prefix" node.

2.5D Anime model: Aux. How to substitite: with anime model of your style choise. Examples of 2.5D anime models: Hassaku, AniWave, DivineEleganceMix. Add universal quality tags recommended by the model author to "Quality prefix" node.

2D Anime model: ObsidianV3-Flat. How to substitute: don't know, really. Put your favorite model of unique style, I guess. Add universal quality tags recommended by the model author to "Quality prefix" node.

VAE to encode realistic: vae-ft-mse-840000-ema-pruned. Standard, recommended.

VAE to decode anime: kl-f8-anime2. How to substitute: vae-ft-mse-840000-ema is fine.

Embeddings: easynegative (recommended), bad-hands-5 (optional).

How to use

Choose where to get image and prompts from and connect nodes for image, positive and negative into ">> Route your inputs here <<" group:
- By default, manual written prompts are used
- Turn on (Ctrl + M) and connect "WD 1.4 Tagger" node to generate prompt from picture
- Turn on (Ctrl + M) and connect "SD Prompt Reader" node to get prompts from metadata
Run it!