Prefect Pony XL · Text to Image

Two Pony nodes ship in the PopNow pack, and they're not the same tool with a checkbox between them. The LoRA one costs 3.6 credits and takes up to three of your LoRA files on an aspect_ratio dropdown. This one costs 2.25 credits - the cheapest image in the pack - takes no LoRAs at all, and asks you for a size in pixels instead of a ratio.

Which is the better node? If you own Pony LoRAs, the other one. If you don't, this is the cheapest Pony render you can buy anywhere, and at 2.25 credits - under two and a half cents a render - it's a hard price to argue with.

Why anyone still wants Pony

The Pony ecosystem doc answers this properly, so the short version: AstraliteHeart trained an SDXL fine-tune on a huge, densely-tagged danbooru/e621 dataset, and the tagging quality bought it prompt comprehension that no other anime model matched. It also destroyed its tolerance for prose. Plan language prompting does not work; tag sequences do.

That means the prompt box here is not a description. It's a list. score_9, score_8_up, score_7_up, 1girl, solo, long hair, school uniform, classroom, window light, depth of field.

Put the score string first, every time. It's not a style flourish, it's the quality switch - the KB explains the training accident that made it mandatory, and trying to use a single score_9 instead of the full run is the most common reason people think Pony "got worse". Then subject, then wardrobe, then lighting and lens, in that order. The tooltip says exactly that.

The inputs

prompt - tags, score preamble first.

size is the interesting one, and it's why this node feels more native than its sibling. It's written as width*height and every value is an SDXL bucket: 1024*1024, 896*1152, 1152*896, 832*1216, 1216*832, 768*1344, 1344*768. Those aren't arbitrary - they're the aspect buckets SDXL was trained on, and using them avoids the weird stretched compositions you get from forcing a 3:2 ratio onto an SDXL grid. 832*1216 is the classic anime portrait, and 1024*1024 defaults are fine for anything square.

output_format defaults to png here, which is the right default for illustration - flat colour and linework are exactly what JPEG smears.

seed starts fixed. Randomize to explore at 2.25 credits a roll; at that price, exploring is genuinely cheap.

Outputs: images and task id.

Install

Manager → search PopNow → Install → restart, or:

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

Then a key from popnow.ai/api/console/keys into Settings > PopNow, or for a headless box:

POPNOW_API_KEY=sk-popnow-... python main.py

No model download, no dependencies, no GPU. Worth knowing for the whole pack: the node list is rebuilt from your account's catalog at every startup, so a restart is also how new PopNow models appear without updating anything.

Where people get burned

Prose prompts. "A girl in a classroom, soft afternoon light, cinematic" will produce something, but it'll be mushy and generic next to a tag string. Pony rewards specificity in tag form. If you're coming from Flux or SDXL-realistic habits, this is the habit to unlearn - the prompt engineering notes cover how differently tag-trained models behave.

Wrong-shaped sizes. The dropdown is deliberately limited to SDXL buckets. Squeezing a panorama out of a Pony checkpoint by picking a wider bucket gets you duplicated subjects, not a wider scene.

No LoRA support here, by design. If you need a character or a style LoRA, use the LoRA variant of this node - this one has no loras socket at all, and there's nothing to plug in.

Rating tags are part of the system. Pony's content rating comes from the same tag vocabulary as its quality score. If you're generating SFW work, say so in tags rather than expecting the model to infer it.