Pony Diffusion XL prompting
Pony Diffusion XL (PDXL) is one of the most capable SDXL fine-tunes, spanning anime, semi-realistic, and furry styles from a single checkpoint — but it has its own non-obvious prompting conventions. Get the score chain and source tag right and it produces clean, controllable images; skip them and output looks flat. This guide assembles the correct prefix and settings for you.
How it works
Pony was trained with an explicit aesthetic-scoring scheme, so it expects a
descending score chain at the very front of the prompt:
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up. This chain
is the heart of PDXL prompting and is essentially mandatory. After it comes a
source tag — source_anime, source_pony, source_furry, or
source_cartoon — which steers the overall style, followed by a rating tag
(this builder centers on rating_safe) and then your subject tags. The tool
builds that prefix for your chosen style and pairs it with a recommended negative
preset and CFG/step settings.
The four source tags and what they do
| Source tag | Visual style |
|---|---|
source_anime | Clean anime linework, flat cel shading, vivid colors |
source_pony | Stylised My Little Pony-adjacent aesthetic |
source_furry | Anthropomorphic characters, furry art community conventions |
source_cartoon | Western cartoon look, thicker outlines, exaggerated shapes |
Omitting the source tag leaves Pony guessing, and results lean toward whichever style was most common in its training data for your subject matter. For predictable output, always include one.
Recommended negative prompt
A solid negative prompt pairing for most PDXL generations:
score_6, score_5, score_4, worst quality, low quality, bad anatomy,
bad hands, text, watermark, blurry, jpeg artifacts, extra limbs
The leading score_6, score_5, score_4 in the negative tells Pony to avoid low-rated training images — this alone eliminates a lot of the muddy, flat look. Standard quality negatives handle anatomy and compression artefacts.
CFG, steps, and sampler
A reliable baseline:
- CFG scale: 6–7 (higher values over-saturate or add halos; lower values go soft)
- Steps: 25–35 (below 20 often looks unfinished; above 40 rarely adds value)
- Sampler: Euler Ancestral or DPM++ 2M Karras for most subjects
Pony handles a wider CFG range than vanilla SDXL, but the sweet spot is still moderate. If skin tones look waxy or colors are neon, lower the CFG by 1–2 before anything else.
Tips and notes
- Never omit the score chain. It’s the difference between crisp and muddy output — always lead with it.
- The source tag is your style switch. Same subject, different
source_*tag, very different look. Experiment to find your aesthetic. - Use the negative score chain too. Pairing positive
score_9...with negativescore_4, score_5, score_6sharpens results noticeably. - Moderate CFG. Around 7 with ~28 steps is a dependable baseline; back off if colors look over-saturated.