ComfyUI on your machine

ComfyUI runs the image model. Orihai talks to it over HTTP and never loads a model itself, so ComfyUI has to be running before you ask for a picture.

Graphics card

An NVIDIA card with 8 GB of memory generates an anime SDXL image in a few seconds. 6 GB works and is slower. AMD works on Windows through ComfyUI's DirectML build, slowly. Without a card it falls back to the processor, which takes minutes per image rather than seconds.

Disk

About 7 GB for ComfyUI and one checkpoint. Each extra checkpoint is another 6 GB.

Download the portable Windows build from the ComfyUI releases page: the file is ComfyUI_windows_portable_nvidia.7z, about 1.5 GB. Extract it anywhere you have room. It carries its own Python, so there is nothing else to install and nothing added to your system.

Inside the extracted folder you get:

run_nvidia_gpu.bat

Starts the server. Leave the window open while you use Orihai. It prints To see the GUI go to: http://127.0.0.1:8188 when it is ready.

ComfyUI/models/checkpoints/

Where image models go.

ComfyUI/models/loras/

Where LoRAs go.

Start it once before downloading anything and open http://127.0.0.1:8188 in a browser to confirm it answers.

A checkpoint is the image model. For anime you want one trained on booru tags, because that is the vocabulary Orihai writes its prompts in: comma-separated tags rather than sentences.

Get them from Civitai or Hugging Face. On Civitai, filter by Checkpoint and sort by downloads. Pick the .safetensors file, never a .ckpt: the safetensors format cannot execute code when it loads.

These are the families worth knowing about, newest first:

Illustrious / NoobAI

The current anime standard. Trained on recent booru data, so it knows character names, clothing and pose tags well. Around 6.5 GB. Start here.

Pony Diffusion V6 XL

Older, still widely used, and unusually good at multiple characters in one image. Wants the quality prefix score_9, score_8_up, score_7_up at the front of the prompt, which goes in the profile's positive template.

Animagine XL

A cleaner, more conventional anime look. Lighter on style, heavier on anatomy.

Put the downloaded file in ComfyUI/models/checkpoints/ and restart ComfyUI. It appears in Orihai's checkpoint dropdown the next time the profile form fetches the list.

You do not download samplers. They are algorithms built into ComfyUI, and the dropdown in the image profile lists the ones your installation has.

Choose euler_ancestral. Orihai sends the normal scheduler with every workflow, and euler_ancestral is the sampler that pairs correctly with it on every anime checkpoint above. It is also what the model authors sample their own preview images with.

dpmpp_2m is the other common choice, and it expects the karras scheduler. Paired with normal it produces a visibly softer, washed-out image, so leave it alone until the app can send a scheduler of its own.

A new ComfyUI profile fills the field with dpmpp_2m by default. Change it.

Do not type k_euler_a or Euler a. Those are the AI Horde and Automatic1111 spellings, and ComfyUI rejects the whole workflow rather than guessing.

Settings, then Images, then a new profile with ComfyUI as the provider.

Endpoint

http://127.0.0.1:8188, unless you started ComfyUI on another port. Once this is set the form fetches the checkpoint and sampler lists and turns both fields into dropdowns.

Checkpoint

From the dropdown. If it is empty, ComfyUI is not running or is on a different port.

Sampler

euler_ancestral, per above.

Steps

28 to 30. Fewer than 20 leaves anime models undercooked; more than 35 changes very little and costs seconds.

CFG scale

5.0 for Illustrious and NoobAI, 7.0 for Pony and Animagine. Higher follows the prompt more literally and burns colours doing it.

Size

1024 by 1024, or 832 by 1216 for portraits. The 512 default belongs to older models and gives an SDXL checkpoint mush.

Hires fix

On, scale 1.5, denoise 0.45. It samples a second pass at a larger size, which fixes the small faces and hands that come out of a single pass. It roughly doubles the time per image.

Negative prompt

lowres, bad anatomy, bad hands, text, error, missing finger, extra digits, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, username, blurry

Positive template

masterpiece, best quality, amazing quality, plus score_9, score_8_up, score_7_up at the front if you chose Pony. Orihai adds the rating, the character and the scene after it.

A LoRA is a small file, 50 to 200 MB, that teaches the checkpoint one character or one style. Download it from Civitai the same way, put it in ComfyUI/models/loras/, and restart ComfyUI.

The LoRA field in the image profile is free text rather than a dropdown, because ComfyUI does not report installed LoRAs in the same call the checkpoints come from. Type the filename as it sits on disk, without the .safetensors. A wrong name fails the whole generation.

Strength runs 0 to 1. Character LoRAs want 0.7 to 0.9; style LoRAs want 0.4 to 0.6, because a style at full strength overrides the checkpoint's own.

ComfyUI rejected the workflow

A checkpoint, sampler or LoRA name it does not have. Restart ComfyUI if you added a file since the profile was saved, then reopen the profile so it fetches the lists again.

The profile will not save

The form refuses a checkpoint or sampler that is not in the fetched lists, so a name ComfyUI would reject never reaches it. Start ComfyUI, then press Retry.

ComfyUI did not finish in time

Generation is capped at two minutes. On a slower card, turn hires fix off or drop the steps.

The picture is of the wrong scene

Orihai's utility model writes the tags for each illustration. If it is not configured, or its key is missing, the app tells you which stage failed rather than illustrating something else. Check the utility model profile under Models.