Prefect Pony XL LoRA · Text to Image

Pony-lineage SDXL is the most-installed thing in this hobby and the one nobody wants to reinstall. This node runs a Pony XL checkpoint - Prefect, a well-known member of the family - on someone else's GPU, edits included: score-tag prompting, nine native SDXL ratios, guidance from 1 to 20, and up to three of your own Pony LoRAs, for $0.024 an image.

The prompt paradigm is the whole game

You can't prompt Pony like Flux. It's a booru model, so it wants comma-separated tags, and its house quality convention is the score-tag prefix - score_9, score_8_up, score_7_up - which isn't decoration. AstraliteHeart hand-ranked roughly 20,000 training images and captioned the dataset with those labels so the model would learn concepts without learning ugliness; a training error fused them into the verbose string everyone types to this day. The pack's own tooltip on this node puts it plainly: score tags first, LoRA trigger words at the start.

If you've been prompting Z-Image or Flux with prose and wondering why Pony answers seem mushy, that's the reason. Tags for content, tags for pose, tags for the character name, plus source_anime-style tags if you want the look steered, and rating tags if you care about that axis. And a warning that travels with the whole family: none of this transfers. Score tags are inert on Illustrious, SDXL and Flux, just as masterpiece, best quality does very little here. The practice is fading in the wider community, but on a Pony checkpoint it still earns its keep.

Where this sits against local: Pony users have spent years accumulating merges, LoRAs and a prompt library. You don't get those - you get a clean hosted checkpoint. What you do get is your LoRAs, attached by URL, and a $0.024 image with no 6GB download, no 24GB-vs-8GB drama and no VAE hunting.

Inputs

• prompt - tags with score tags first; trigger words at the front of the line.
• aspect_ratio - default 1:1, nine options at native SDXL geometry. Pick one deliberately: cropping a 1:1 render to 16:9 costs you the composition you paid for.
• guidance_scale - 1 to 20, default 3.5, and a value worth noticing: Pony users habitually type 5 to 7 locally. If results ignore parts of your prompt, raise it a notch; if they go crunchy, bring it back down.
• output_format - jpeg by default; PNG and WebP available. Use PNG if you're going to feed this into an upscale.
• seed - same seed and settings, same result.
• loras (optional) - the loras output of a SpicyAPI LoRA node, up to three, each with its own strength. It must be a Pony-family SDXL LoRA: the pack's tooltip is explicit about that, and it's not a technicality - a Flux or Illustrious LoRA here will either do nothing or wreck the render. The URL has to be a direct .safetensors link that opens without a login.

Outputs are images and task id, the latter for support.

Installing it

ComfyUI Manager → search SpicyAPI → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart ComfyUI, open Settings > SpicyAPI, paste a key from spicyapi.ai/console/keys, and restart once so your node list is built from your account. The pack installs no Python dependencies and downloads no weights - your machine only needs to reach api.spicyapi.ai. Server users: SPICY_API_KEY.

What to watch

At $0.024 an image, a hundred-image booru-tag sweep is $2.40, which is cheap until you're on your eighth sweep of the evening. Set Max cost per run in Settings > SpicyAPI; the node quotes each job and refuses anything above your ceiling before creating a task. Seeds default to fixed on purpose, so re-queueing the same graph reuses the first image rather than quietly selling you a second - switch control after generate to randomize for new takes.

Your prompts and any LoRA links go to a server on every run; that's the mechanism. This variant is described as unfiltered on prompts, weights and results, but it's still a hosted service, and a task that returns content_rejected means the model declined - typically without charging you. Cancelling a run stops your wait, not the job.