Skip to main content

Uncensored — This model declares no content filter. SpicyAPI forwards the request without adding a platform content scan; legal and consent duties remain with the customer.Browse every model at this filtering level
prefect/pony-xl/text-to-image

Prefect Pony XL is a community anime and illustration checkpoint from the Pony XL lineage, built on SDXL. It takes comma-separated tag prompts of up to 5,000 characters, seven native sizes and an optional seed, in JPEG, PNG or WebP.

From $0.015/image
Run model

Workbench

Input

Text to image
179/5000
Tag-style prompting steers this checkpoint far harder than prose: comma-separated tags, quality tags first, then subject and wardrobe, then lighting and lens. Up to 5,000 characters. at least 1 characters · at most 5000 characters
Seven native SDXL shapes. Arbitrary width and height are not accepted — off-ratio sizes on an SDXL backbone are where anatomy tends to break, so only the trained pixel counts are exposed.
Output format
PNGJPEGWebP
PNG is lossless and suits a hand-off into retouching, WebP suits a frame going straight onto a page, JPEG suits the smallest transfer. The choice does not affect the generation or the price.
Seed
Pass a seed to reproduce a frame exactly with the same prompt and size; omit it and every call explores. The price is the same either way. no lower than 0 · no higher than 2147483647
Safety checkerOff

SpicyAPI does not run a safety checker on this model. Your prompt and media go to the model as written, and the model itself decides what comes back.

Checking your account before enabling submission and pricing.

Output

Ready for a new run
example-silver-knightExample result
Animate
Edit
Split into layers

Related models and next tasks

Missing the model you need?

Tell us what to add and where you would use it. Requests go to the people who decide what gets onboarded next.

Tag-prompt anime art in four fields

Prefect Pony XL is a community anime checkpoint from the Pony XL lineage, built on the SDXL backbone and served here as a single text-to-image endpoint. One request carries four fields: a prompt of up to 5,000 characters, one of seven native SDXL sizes, an optional seed and an output format of png, jpeg or webp. It reads comma-separated tag strings the way local SDXL checkpoints do, quality tags first and subject behind them, and it returns one image per call.

Open the playground
Request fields
Four: prompt, size, seed, output_format
Prompt length
1 to 5,000 characters, the only required field
Sizes
Seven native SDXL shapes, 1024*1024 by default
Seed
Optional integer, 0 to 2,147,483,647
Output format
png (default), jpeg or webp
Per call
One image, asynchronous, 1,800-second ceiling
Lineage
A Pony XL checkpoint merge on the SDXL backboneTag-style prompting behaves here the way it behaves in local SDXL tooling, because that is the lineage this checkpoint comes from. It is a merge tuned for anime and illustration, which is also its boundary: photoreal briefs come back stylised, so photographic work belongs on a photoreal model in this catalogue instead.
Interface
Four fields, and the schema refuses everything elseprompt, size, seed and output_format are the entire request body. Any other key fails validation rather than being quietly dropped, so a call that returns an image is a call that used every value you sent. There is no image input, no negative prompt, no step or guidance control, no LoRA slot and no batch count.
Content policy
No maturity flag, because there is nothing to switch onThe four fields are the whole request: no per-request declaration, no uncensored twin of this endpoint and no content filter of ours between the prompt and the checkpoint. You remain responsible for the lawfulness of what you generate and publish, and the underlying checkpoint ships under the CreativeML Open RAIL++-M licence, whose use restrictions travel with the images it produces.

What the Prefect Pony XL API does well

  • Tag strings from a local setup paste in unchanged

    The prompt dialect does not change on the way to an API: comma-separated tags, quality tags in front, then subject and wardrobe, then lighting and lens, inside 5,000 characters. Every sample prompt further down this page opens that way. A tag library built up over months of local runs keeps working; the one habit that does not carry over is the separate negative field, because everything here goes into the single prompt.

  • Seven native SDXL sizes, picked before generation

    1024*1024, 896*1152, 1152*896, 832*1216, 1216*832, 768*1344 and 1344*768, with 1024*1024 applied when the field is left out. Arbitrary width and height are not accepted: off-ratio requests are where faces, hands and armour tend to fall apart on this architecture, so only the trained shapes are exposed and any other ratio is a crop on your side.

  • A seed that rebuilds the exact frame

    Pass an integer between 0 and 2,147,483,647, and the same seed with the same prompt and the same size returns the same image. Leave it out and every call explores a fresh frame, which is the right default while the wording is still moving. It is the only grip on a composition here, since this endpoint takes no image input.

  • Three output formats, none of which touch the picture

    png is lossless and suits a hand-off into retouching, webp suits a frame going straight onto a page, jpeg suits the smallest transfer. The choice changes the delivered file alone: the same prompt, size and seed produce the same image in all three, and the per-image charge is identical as well.

  • One image per call, one line on the bill

    No count or variant field exists that could multiply a single call, so a set of four frames is four calls, each with its own seed and its own charge. That is also what makes a bad frame cheap to fix: it is rerun on its own, rather than by regenerating the whole set around it.

  • Asynchronous, with a stated ceiling

    The create call returns a task rather than an image; the frame is read from that task once it reaches a terminal state, or from the callback you registered with the request. 1,800 seconds is the point at which a task is cut off, not an expected wait, and because tasks are independent a queue of prompts runs in parallel from one worker.

Who Prefect Pony XL is for

  • Illustrators moving off a local rig

    The same prompts, without the checkpoint file underneath

    The tag conventions transfer directly, so the move is mostly a change of transport: what you stop maintaining is the GPU and the model file. The trade is stated in the schema rather than hidden — sampler settings are fixed upstream, there is no LoRA slot, and steps and guidance are not exposed.

  • Character and reference sheets

    One character across the seven shapes

    Hold the seed and change only the framing tags: 832*1216 for a full-body sheet, 1024*1024 for an avatar crop, 1344*768 for a banner. The frame hangs on the size as well as the seed, so switching size re-composes the picture instead of re-framing the same one — worth knowing before a sheet is planned around it.

  • Visual-novel and game teams

    Assets at the aspect they ship at

    832*1216 and 896*1152 match the sprite and CG frames most visual-novel engines expect, so a portrait arrives at the shape it will be used at instead of being cropped out of a square. Generating at the destination shape also keeps the staging the model composed for you.

  • Uncensored illustration

    The ordinary call, with no review queue in front of it

    Briefs at every content level run on this same endpoint, with no flag to set and no second endpoint to call. What replaces the toggle is your own judgement: you are responsible for the lawfulness of what you generate and publish, and the licence on the underlying checkpoint carries use restrictions that follow the output.

How to call the Prefect Pony XL API

  1. Settle the size before you write the prompt

    Composition follows the frame, so pick one of the seven values first; omit the field and you get 1024*1024. A size outside the list is refused at request time rather than rounded to the nearest shape, and a ratio that is not on the list has to be cropped on your side afterwards.

  2. Write the prompt as a tag string

    Quality tags first, then subject and wardrobe, then lighting and lens, comma-separated and inside 5,000 characters. Prose is accepted, it just steers more weakly. There is no negative prompt on this endpoint, so what you want kept out of the frame is steered by the tags you do write, not by a second field.

  3. Send a seed once a frame might need rebuilding

    Omit it while you are still exploring and every call returns a different frame. Once a result is worth keeping, store the seed next to the prompt and the size — all three are needed to rebuild it — and pull the file down, because generated media is retained for 14 days and then deleted.

  4. Submit, then poll the task to a terminal state

    The call hands back a task, not an image. Poll it until it settles, or take the callback you registered, and allow for the 1,800-second ceiling before a task is cut off. One task yields one image, so four frames are four tasks, and they can run in parallel on separate seeds.

  5. Change one variable per iteration

    On a held seed, swapping a single tag usually keeps the staging and moves the detail, while rewriting the prompt wholesale returns a different picture rather than a variation of it. Switching the output format changes nothing about the image; switching the size re-composes it from scratch, even on the same seed.

Prefect Pony XL worked examples

  • A wide night-street character shot

    The landscape shape is asked for in the request rather than cropped in afterwards, and the neon is carried by tags because there is no lighting setting to reach for. size is 1216*832.

    score_9, score_8_up, 1girl, solo, silver twin tails, oversized varsity jacket, denim shorts, standing on a neon-lit city street at night, rain reflections, looking at viewer, cinematic lighting, detailed background

  • A full-body character sheet

    832*1216 is the portrait shape these are usually generated at. Armour detail is where SDXL checkpoints tend to fall apart at off-ratio sizes, which is the argument for a native shape rather than a resize.

    score_9, score_8_up, 1girl, solo, long white hair, ornate silver armor, red cape, holding a longsword, standing on castle ramparts at dawn, dramatic sky, full body, detailed armor

  • Two figures in one frame

    A scene rather than a portrait, at 1344*768, the widest shape available. Pin the seed on the version that holds together, then iterate the wardrobe and the light around it one tag at a time.

    score_9, score_8_up, 2girls, red and white miko outfits, sweeping fallen maple leaves at a shrine, autumn morning mist, stone lanterns, wide shot, detailed background

Prefect Pony XL API questions

Write for the checkpoint rather than for a parameter sheet: a tag string that names quality, subject, wardrobe and light in that order steers this model harder than any setting would, and there are no settings here to reach for. The seed and the size are the only other things holding a frame in place, so both belong in whatever record you keep next to a prompt. A miss is answered by changing one tag on the same seed; a keeper is answered by writing the seed and the size down and pulling the file before the 14-day retention window closes.

All Prefect Pony XL models2