Setup: ComfyUI

TomoriBot can generate images and videos through your own ComfyUI instance. It drives ComfyUI by submitting an API-format workflow with your prompt/size substituted in, then polls ComfyUI’s /history endpoint until the output is ready.

This guide covers installing/running ComfyUI and registering it. For authoring or editing a TomoriBot-compatible workflow (the {TOMORI_*} placeholders), use the in-Discord deep dive, open /help, choose Features, then Custom Endpoints, and use the workflow README on GitHub.

No env vars needed

ComfyUI is registered through Discord slash commands and stored encrypted in the database. See the local endpoints hub.

Hardware requirements

Image and video generation is GPU-bound: the heavy cost is VRAM (your graphics card’s memory, separate from system RAM), set by the model checkpoint your workflow loads, not by ComfyUI itself. An NVIDIA GPU is strongly recommended. The two workflows TomoriBot ships are both modern, heavier-than-SDXL models:

Bundled workflow	Base model	Practical VRAM	Notes
Anima v1 (image)	Qwen-Image (~20B), fp8	~16 GB floor · 24 GB comfortable	Text encoder + VAE add ~8-10 GB of overhead. Below 16 GB, use a GGUF build + --lowvram.
WAN i2v loop (video)	Wan 2.2 14B, fp8 + 4-step LightX2V LoRAs	~16 GB workable · 24 GB+ comfortable	Heaviest option expect minutes per clip. Offload the UMT5 text encoder to RAM (t5_cpu, needs 24 GB+ system RAM) on smaller cards.

Both bundled checkpoints are already fp8-quantized to fit consumer cards. If you have less VRAM, swap the UNET for a smaller quant and enable ComfyUI’s --lowvram / CPU offload. CPU-only diffusion is impractical (many minutes per image, worse for video) and can exceed TomoriBot’s poll window, so a GPU is effectively required for regular use.

Going lower: pick a GGUF quant

Quantization stores each model weight in fewer bits to cut VRAM and disk use, at a small precision cost. The bundled fp8 files are a mild form of it, and to shrink further, download a GGUF build of the model from Hugging Face: the code in names like Q4_K_M / Q5_K_M is the bits-per-weight, and 4-bit (Q4) or 5-bit (Q5) is the usual sweet spot with most of the quality for a fraction of the fp8/fp16 footprint. Below 4-bit shrinks more but degrades quickly. Loading GGUF UNETs in ComfyUI needs the ComfyUI-GGUF custom node.

1. Run ComfyUI with the API enabled

Install ComfyUI per its README and start it so it listens on the network:

python main.py --listen 0.0.0.0 --port 8188

--listen 0.0.0.0 matters if TomoriBot runs in Docker or on a different machine: the default binds to loopback only. Confirm reachability from the machine the bot runs on:

curl http://127.0.0.1:8188/system_stats

If you want to test, load the model checkpoints your chosen workflow expects and do one manual generation in the ComfyUI web UI to confirm it works end to end before wiring up TomoriBot.

2. Grab a TomoriBot workflow

Download a ready-to-use API-format workflow. Examples can be found in the repo under assets/comfyui-workflows/:

Workflow	Modes
Anima v1 (image) : tomoribot-anima-v1-comfyui.json	txt2img, img2img, inpaint
WAN i2v loop (video) : tomoribot-wan-i2v-loop-video.json	image-to-video

These are API format (the JSON ComfyUI exports via Save (API Format)), not the regular UI-save format. If you author your own, it must contain the {TOMORI_*} placeholders TomoriBot substitutes (prompt, width/height, seed, reference images, etc.). See the workflow README and the Custom Endpoints page under Providers in /help.

3. Register it in Discord

Run /providers (or /personal providers), choose Add New Custom Endpoint, and enter:

Field	Value for ComfyUI
endpoint_label	A name you choose, e.g. home-comfy
API Compatibility	ComfyUI
endpoint_url	http://127.0.0.1:8188 (root, no /v1)
auth_token	(leave blank unless your ComfyUI is behind auth)

After saving the connection, select it and use its model dropdown to add an Image or Video model. Enter the checkpoint’s exact code name and upload the workflow .json you downloaded from Step 2. The model capability must match the workflow (image workflow → image, video workflow → video).

An image model also asks for its Image Capabilities: text to image, reference image, inpainting, and negative prompt. Tick only the modes your workflow actually implements, because Tomori offers the tool only the modes you declare. Inpainting appears for ComfyUI connections only, since no other API compatibility accepts a mask. Editing the model later reopens the form with your current selection, so changing a code name will not clear it.

Adding the model makes it the active image/video model automatically. Trigger generation by asking Tomori directly in chat. If it isn’t active for some reason, run /config > Models > Switch Models and select your registered ComfyUI endpoint.

Troubleshooting

• Unreachable on add: ComfyUI bound to loopback while the bot is in Docker / on another host. Start it with --listen 0.0.0.0 and use http://host.docker.internal:8188 or the LAN IP.
• Generation never completes: TomoriBot polls /history until the output appears. Cold starts and large models on CPU can exceed the poll window.
• Prompt/size ignored or wrong output size: the workflow is missing the required {TOMORI_*} placeholders, or you uploaded a UI-format export instead of API format.
• Wrong capability: an image workflow registered under video (or vice-versa) won’t run. Re-add it under the matching capability.