Character art with ComfyUI

There are two ways in, and the choice comes down to whether you have a GPU.

• Comfy Desktop runs on your own machine. It's free, everything stays local, and there's no per-image cost — but you want a reasonably recent GPU with 8GB of VRAM or more, and the first setup takes a while because you're downloading multi-gigabyte models.
• Comfy Cloud runs in the browser on their hardware. Nothing to install, works on a laptop, and it's the sane option if you don't have a GPU — it's the paid side of the product.

If you're unsure: try Desktop first. If the models are painfully slow or won't fit in memory, that's your answer.

1. 1

   Install and open a template

   ComfyUI ships with example workflows. Don't build a graph from scratch on day one — open the basic text-to-image template and run it once so you know the install works.

2. 2

   Get a checkpoint that suits your character

   The checkpoint is the model, and it decides the look far more than your prompt does. An anime-styled character and a photoreal one want completely different checkpoints; fighting a model to produce a style it wasn't trained for is the most common reason people give up. Use ComfyUI's model manager rather than dropping files into folders by hand.

3. 3

   Set the resolution to 3:4 before anything else

   ShikuAI cards are 3:4 portrait, and uploads are resized to 800 × 1066 with anything outside that shape cropped away. Generating at the wrong ratio is how you end up with a character whose head is cut off on their own card.

   864 × 1152 is the number to use: exactly 3:4, about one megapixel, and a size SDXL-family models are comfortable with. Then let the upload downscale it.

4. 4

   Write the prompt in layers

   Subject first, then appearance, then framing, then style — for example: who they are, hair and eyes and clothing, upper body portrait, looking at viewer, and finally the rendering style. Keep the framing terms in there; without them you get full-body shots where the face is fifty pixels tall and unusable at card size.

5. 5

   Upscale, then upload

   Add an upscale pass at the end of the graph. It costs a few seconds and it's the difference between a card that looks sharp and one that looks like a thumbnail. Then drop the file into the artwork slot on the character creator.

This is the part that matters most here and that generic tutorials skip. A character on ShikuAI is a person people come back to, so their pictures need to look like the same person.

• Fix the seed. Once you get a face you like, lock its seed and change only the prompt. Random seeds give you a new person every time.
• Keep the description verbatim. Copy the exact appearance clause between generations rather than rewriting it. Small wording changes move the face more than you'd expect.
• Save the workflow. ComfyUI embeds the whole graph in the PNG it produces — drag that image back into ComfyUI and you get the settings back. Keep one good render per character as your reference file.
• Fix faces with inpainting, not by rerolling the whole image. Mask the face, regenerate just that region, keep the composition you already liked.

Generating the art doesn't change the rules it has to follow. Anything you upload still has to satisfy ShikuAI's content guidelines — in particular, no sexual content involving anyone who reads as a minor, and no depictions of real, identifiable people. A model will happily produce both if you ask; that is your responsibility, not the tool's.