Pony Diffusion XL (custom model)

This is not a model node - it's a doorway node. RunwareArch_pony lives in the pack's Runware/Custom models folder, and its job is to let you run any community Pony-architecture checkpoint from the Runware catalog, without downloading a single file. You type (or search) a model AIR into the model field, and the node runs that exact checkpoint in the cloud. Pony Diffusion V6 XL is the SDXL fine-tune that owned 2024's stylized and NSFW generation - famous for anatomy, niche fetishes, and that botched score-tag system you had to paste into every prompt. Hundreds of merges (AutismMix, PonyRealism, and friends) made "Pony" a family, and this node is how you browse that family from inside ComfyUI.

Why would you use this instead of the pack's fixed image nodes? Flexibility. The dedicated nodes are pinned to one model; this one can point at any Pony checkpoint Runware hosts - including the merge you actually want for a specific look. If you already run Pony locally, this is also a nice escape hatch for when you're away from your GPU and just need a quick render.

What you set

Required: model (an AIR string like runware:257749@290640 - use the search button to browse the catalog), positivePrompt, width, height (128–2048, step 8).

Then the full SDXL feature set you'd expect:

• negativePrompt - works here, unlike on the distilled models. Pony is classic guidance, so both prompts matter.
• CFGScale toggle + CFGScale_value - Pony's sweet spot is around 4-6, lower than SD 1.5's 7+. The default is fine; you're mostly nudging, not launching.
• steps toggle + steps_value - 20-30 is the Pony norm.
• clipSkip - the classic anime trick, 1-2 layers, because Pony's CLIP is overtrained on tags.
• vae - override the base VAE if a specific checkpoint wants one.
• promptWeighting - compel or sdEmbeds syntax for prompt weighting.
• Sockets: lora, controlNet, ipAdapters, embeddings, refiner, outpaint, maskImage, seedImage - the full builder-node spread.
• scheduler, seed, numberResults (up to 20), outputFormat, outputQuality, and advanced_json (tooltip: for hiresFix).

Output is image (IMAGE).

How it works

Same pack machinery as everything else - the request carries the model AIR you chose, plus taskType, over REST via the Runware SDK, and the finished IMAGE comes back. The subtle part: because Pony checkpoints live in the catalog with known architectures, Runware picks the correct sampler settings for the architecture automatically. You still tune CFG and steps, but you don't fight the plumbing. The title bar shows cost, and Pony at 1024 is cheap enough that trying three merges to find your look is an afternoon hobby, not a budget crisis.

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 - the catalog hosts them. Dependencies are runware-sdk, pillow, soundfile. API key via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

The model AIR is the failure point - a typo or stale AIR means a failed request, so use the search button. If your Pony output ignores the negative prompt, check CFGScale isn't ≤ 1 (negative is ignored there by design). And if you expected Pony's score-tag behavior, remember you still need those tags in the prompt - the node doesn't add them for you.