NoobAI XL (custom model)

NoobAI XL is the community fork that finished what Illustrious started - trained further on Danbooru plus e621 data, and famous for one thing: a vpred variant that fixed anime lighting. Dark scenes, dramatic rim light, night shots that don't look washed out - that's NoobAI's party trick, and it made the model the community's default anime base once Illustrious went paywalled. The problem has always been that vpred is finicky to run: it wants a specific sampler setup (CFG++ around 1.5-2.0, vpred-aware LoRAs, and on ComfyUI a ModelSamplingDiscrete node), and a wrong config silently ruins your lighting - the exact thing you switched for.

This is where RunwareArch_noobai earns its keep. It's the custom-model doorway to NoobAI in the Runware pack: type or search an AIR into the model field and the checkpoint runs in the cloud. The vpred plumbing - model sampling patches, correct sigma handling - is handled by Runware's side, not your graph. You skip the #1 way people ruin NoobAI output.

What you set

Required: model (default civitai:874216@1434498), positivePrompt, width, height (128–2048, step 8).

• CFGScale toggle + CFGScale_value - this is where NoobAI differs from Pony/Illustrious: vpred likes low CFG, around 1.5-2.0, and many users run CFG++ style sampling. If the default doesn't match your favorite local settings, nudge down, not up.
• clipSkip - tag-model standard, 1-2.
• vae - vpred checkpoints sometimes want a specific VAE; this field overrides the base.
• promptWeighting - compel or sdEmbeds.
• Sockets: lora (vpred-aware LoRAs matter here), controlNet, ipAdapters, embeddings, refiner, seedImage, maskImage, outpaint.
• negativePrompt, steps, scheduler, seed, numberResults (up to 20), outputFormat, outputQuality, advanced_json (hiresFix).

Output is image (IMAGE).

How it works

Standard Runware plumbing - request with the model AIR + taskType over REST, IMAGE back, cost on the title bar. The value-add is the same as every custom-model node: the architecture is known to the catalog, so the sampler handling is automatic. For NoobAI specifically, "automatic" is the whole point, because the manual vpred setup is where local users lose their evenings. Swap between eps and vpred checkpoints by changing the AIR and you're done - no graph surgery.

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 fail requests - use the search button. If your lighting comes out flat despite using NoobAI, check you're actually on a vpred checkpoint (some catalog entries are the eps variant) and that CFGScale is in the low 1.5-2 range, not 7. And remember vpred-trained LoRAs on an eps checkpoint is a mix that quietly breaks - keep the variant consistent.