Prefect Pony XL LoRA · Text to Image

Pony XL is the one anime checkpoint the community never replaced, and the reason is unglamorous: nobody else tagged 2.6 million images that carefully. It understands booru tags and nothing else, which makes it either exactly what you want or completely useless depending on how you prompt.

Prefect Pony XL LoRA · Text to Image is a hosted Pony XL derivative at 3.6 credits an image, with a loras input that accepts up to three of your own Pony-family LoRA files. If you've been generating Pony art on a 8 GB card at 832x1216 and waiting, this is the obvious trade: three and a half cents and no fans spinning.

Prompt it like Pony, or it won't work

This is the part that trips people up, and the Pony ecosystem doc has the full story. The short version: Pony was trained with a quality rating system, and a mistake during training welded the specific string of score tags into a single quality signal. You can't use one of them. You have to use the whole run:

score_9, score_8_up, score_7_up, 1girl, solo, ...

The tooltip says it exactly: comma-separated tags with score tags first. Put LoRA trigger words at the very start if you're using a LoRA - before the score tags - because that's how the LoRA saw them in training.

And note what this is not: prose. "A young woman standing in a sunlit meadow with soft light" gets you a worse image than 1girl, solo, meadow, sunlight, depth of field. Pony's understanding of plain language was arguably destroyed by its dataset's density of tags. Write tags.

The inputs

prompt - tag string, score preamble, triggers first. Up to you to get the order right; there's no auto-magic here.

aspect_ratio defaults to 1:1. Nine options, the usual portrait/landscape set plus 21:9 and 9:21 ultrawides. Pony is an SDXL model, so the ratios map onto SDXL-ish pixel counts; portrait 9:16 is where it looks most like the work people post.

guidance_scale defaults to 3.5 - leave it. It's the CFG equivalent, and on a Pony checkpoint the useful band is narrow. If your image is flat and lifeless, raise it a little; if it's crunchy and over-saturated, lower it. Either way, change one thing at a time and remember each test is a billed run.

output_format defaults to jpeg. Switch to png for anything going into another pass.

seed starts fixed. Randomize to explore, at 3.6 credits per take.

loras (optional) is where the node earns its name: up to three LoRA weight files trained for Pony-family SDXL checkpoints, each with its own strength.

Outputs: images and task id.

Feeding it your own LoRAs

Add a PopNow LoRA node, set source and scale, and wire it into loras. source accepts a .safetensors file in ComfyUI/models/loras (uploaded once, reused for twelve hours), a Hugging Face repo id or page, a Civitai page, or any direct https link. For gated downloads, put HF_TOKEN or CIVITAI_API_KEY in ComfyUI's environment - the pack refuses to take tokens from a widget, because widget values end up inside every workflow and every PNG you share. That's a good decision and you should respect it.

Chain three PopNow LoRA nodes to stack three LoRAs. Most LoRA models accept up to three; this one says up to three too.

Install

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

Or Manager → search PopNow → Install → restart. Key from popnow.ai/api/console/keys into Settings > PopNow, or POPNOW_API_KEY=sk-popnow-... before launching. No dependencies, no checkpoints downloaded - your LoRAs are uploaded from disk instead.

Where people get burned

Wrong-family LoRAs. A Pony LoRA on a Pony model, yes. An SDXL-realistic LoRA trained on a photoreal base will fight the score tags and give you plastic. The LoRA training doc is worth reading on what a LoRA actually absorbs.

Trigger words buried mid-prompt lose to the score tags. Front of the string, always.

"No artist tags" is a real Pony limitation - its style comes from the model and your LoRAs, not from artist names. If you want a specific look, that's what the three LoRA slots are for.