Illustrious XL (custom model)

Illustrious XL is the anime model that made Danbooru tags the precision interface for anime generation - characters, poses, outfits, and artist styles resolve to exact tags, and anything with roughly 200+ Danbooru works is baked in. It's an SDXL fine-tune, and like its cousin NoobAI it's the kind of model where the community's knowledge lives in the prompt vocabulary, not in sliders. RunwareArch_illustrious is the pack's custom-model doorway to it: type or search an AIR into the model field and the checkpoint runs on Runware's hardware, no 7GB download, no VRAM check.

Worth knowing the family drama, because it explains why you'd pick this over the Pony node: Illustrious v0.1 was deliberately released under-trained so the community could finish it (they did - WAI, Hassaku, and the NoobAI fork all descend from it), then the developer paywalled v1.0 and retroactively changed v0.1's TOS. The family thrives through v0.1 fine-tunes, and this node lets you hop between them by swapping one AIR string. If your workflow runs on a specific Illustrious merge, this is the node that keeps it alive in the cloud.

What you set

Required: model (default civitai:140272@1617798 - use the search button to browse), positivePrompt, width, height (128–2048, step 8).

The settings are the full SDXL spread, with the anime-relevant ones front and center:

• CFGScale toggle + CFGScale_value - Illustrious likes around 4-6, a bit under Pony. Lower is safer than higher.
• clipSkip - the tag-model classic, usually 1-2. Illustrious's CLIP lives and breathes tags.
• promptWeighting - compel or sdEmbeds.
• vae - override for checkpoints that want a specific VAE.
• Sockets: lora, controlNet, ipAdapters, embeddings, refiner, outpaint, seedImage, maskImage.
• negativePrompt, steps (20-30 norm), scheduler, seed, numberResults (up to 20), outputFormat, outputQuality, advanced_json (hiresFix).

Output is image (IMAGE).

How it works

Identical plumbing to every Runware node: request with the model AIR + taskType over REST via the SDK, IMAGE downloaded back, cost on the title bar. Because Illustrious is SDXL architecture, the sampler and VAE handling are automatic - your job is the tags. And they are the job: on this model, the difference between a good image and a great one is whether you wrote score_9, score_8_up-style rating tags and proper Danbooru syntax, not whether you moved CFG by 0.5.

Installing

ComfyUI Manager → search Runware → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

No model downloads; dependencies are runware-sdk, pillow, soundfile. API key via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

AIR typos are the #1 failure - search, don't type from memory. If output looks underbaked, remember v0.1 was shipped under-trained by design: that's what fine-tunes fix, so switch AIR to a WAI or Hassaku merge rather than fighting the base. And negative prompt behaving oddly at low CFG is by design (ignored at CFGScale ≤ 1).