PonyXL is the community-favorite anime-focused SDXL fine-tune. Score tags (score_9, score_8_up) drive quality, character-tag vocabulary controls style. Use it when you want anime, fantasy illustration, chibi, mecha, or stylized character work.
PonyXL is the community-favorite anime fine-tune of SDXL — same architecture, retrained on a curated dataset with a unique 'score' tag system that drives quality. Lead every prompt with score_9, score_8_up, score_7_up and the model produces top-tier output; skip them and quality drops dramatically.
Designed for anime, chibi, fantasy illustration, mecha, character art. Reads booru-style tag prompts (the conventions used on anime tag databases) cleanly. Less suited to photoreal; absolutely the right tool for stylized character work.
Score tags
Quality control system
Anime
Primary genre
SDXL
Base architecture
PonyXL's defining trait is its score tag system. Lead every prompt with 'score_9, score_8_up, score_7_up' (in that order, comma-separated) to unlock top-tier output. Without these tags, quality drops dramatically — they're not optional.
PonyXL is purpose-tuned for anime character work — clean line art, expressive eyes, soft cel-shading, dynamic posing. Fantasy illustration with anime sensibility (knights, magic, mythical creatures) also reads natively without engineering the look in prompt.
PonyXL handles non-standard stylized work too — chibi (super-deformed cute characters), storybook illustration, mascot art, kawaii aesthetics. Commits cleanly to non-anime stylized when prompted with explicit medium language.
Particularly strong on mecha — humanoid robots, futuristic vehicles, sci-fi armor with anime sensibility. The base SDXL architecture handles mechanical detail well; PonyXL's fine-tune adds the anime-specific composition behaviors that make mecha read intentional.
Each example shows the exact prompt that produced the result. Copy any prompt with one click.
3:4 · score_9 lead · 1 credit
score_9, score_8_up, score_7_up, anime style portrait of a young woman with long pink hair and emerald eyes, wearing a school uniform with red ribbon, soft natural lighting, cherry blossoms drifting in background, masterpiece, best quality
"score_9, score_8_up, score_7_up" leads. "masterpiece, best quality" closes. In between, list visual elements as comma-separated tags. This is the canonical PonyXL prompt shape — adapt it to your subject and the model produces consistent anime output.
16:9 · score_9 lead · 1 credit
score_9, score_8_up, anime fantasy illustration of a knight in ornate silver armor standing in a moonlit ancient forest, glowing magical runes floating around the blade of his sword, dramatic lighting, detailed
Fantasy + anime is PonyXL's sweet spot. Pair score tags with explicit setting (moonlit forest), specific costume detail (ornate silver armor), and one magical element (glowing runes). Less is more — list 4-6 key elements rather than padding with adjectives.
1:1 · score_9 lead · 1 credit
score_9, score_8_up, chibi-style illustration of a small cat wizard wearing a tiny pointed blue hat, casting a glittering rainbow spell, soft pastel colors, cute storybook aesthetic
"chibi-style" + "cute storybook aesthetic" anchors the model into the chibi register. Score tags still required even for chibi — they're the quality gate, not the style gate. "soft pastel colors" anchors the palette.
16:9 · score_9 lead · 1 credit
score_9, score_8_up, anime mecha illustration of a sleek humanoid robot in a futuristic city at sunset, glowing engine vents, cinematic composition, detailed mechanical parts
"anime mecha illustration" + "detailed mechanical parts" + "cinematic composition" — PonyXL renders mecha with both anime line art and mechanical detail cleanly. "Glowing engine vents" gives one lighting element to anchor the time of day.
"score_9, score_8_up, score_7_up" in that exact order, comma-separated, at the start of every prompt. This is non-negotiable for PonyXL — without these tags quality drops dramatically. They are the quality gate, not the style gate.
PonyXL reads booru-style tag prompts (comma-separated visual elements) better than prose. "long pink hair, emerald eyes, school uniform, red ribbon" beats "a girl with pink hair and green eyes wearing a school uniform with a red ribbon".
"masterpiece, best quality, detailed" at the end of the prompt reinforces output quality. Use sparingly — over-stuffing quality tags can compress visual specificity. Two or three at the end is the sweet spot.
PonyXL is anime-tuned. It can produce semi-realistic output but it's not the right tool for photography, product shots, or documentary realism. For those, use FLUX, Seedream, or other photoreal-first models. PonyXL excels on stylized.
PonyXL responds best to 8-15 visual element tags, not 40-50. Over-tagged prompts produce muddled output where the model tries to fit too much detail into one composition. Pick the essential elements; let the model fill in the rest.
3:4 for character portraits. 16:9 for landscape / action / mecha. 1:1 for chibi / icon / square social. 9:16 for vertical full-body character. Generate at the right aspect — PonyXL composes differently per ratio.
| Setting | Values | Notes |
|---|---|---|
| Quality tags | score_9, score_8_up, score_7_up | Required for top-tier output. Lead with these in every prompt. |
| Aspect ratio | 1:1 · 3:4 · 4:3 · 9:16 · 16:9 · 2:3 · 3:2 | Standard aspect coverage. 3:4 for portrait, 16:9 for landscape/mecha. |
| Batch size | 1–4 per call | Each output counts as its own credit charge. |
| Negative prompt | String (optional) | "low quality, blurry, distorted, watermark, signature" is a sane default. |
| Seed | Integer (optional) | Pin for reproducibility across re-runs. |
| Pricing | 1 credit per image | Flat pricing across all aspects. |
PonyXL's training dataset uses a quality-tier scoring system, and the model learned to associate the score_9 / score_8_up / score_7_up tags with the highest-quality outputs. At inference time, leading with these tags steers the model toward top-tier output. Without them, you get below-median quality. They're a critical inference detail, not an optional tag.