Pony Diffusion V6 XL Download — Civitai Files, Setup, and License

Before you download — three things that break Pony for first-timers

Pony V6 XL is a fine-tune of SDXL 1.0 by PurpleSmartAI / AstraliteHeart, released January 7, 2024. The file is straightforward — 6.46 GB, fp16 pruned, .safetensors. The traps are everywhere else:

• CLIP skip must be 2. Author warning, verbatim: "Make sure you load this model with clip skip 2 (or -2 in some software), otherwise you will be getting low quality blobs." This is the #1 reason people think Pony is broken on first run.
• Every prompt starts with the full score string. Use score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up — not score_9 alone. Per the author, the short version "has a much weaker effect" because of how the model was trained. Add rating_safe, rating_questionable, or rating_explicit to control content level.
• The license restricts monetized inference. Pony V6 XL is released under a modified Fair AI Public License 1.0-SD that bans paid generators, paid tiers, and paid apps unless the author grants permission. Civitai and Hugging Face have explicit grants. Personal and research use are fine.

Get those three right and Pony works. Skip them and your first 50 generations will look like watercolor mush.

Quick Reference

Civitai needs a free account for NSFW-eligible downloads. The Hugging Face mirrors don't.

The base Pony V6 setup (download in this order)

You only need three files to get a working Pony V6 install. The checkpoint, the VAE, and the negative embedding. The LoRAs come later.

1. Pony Diffusion V6 XL — V6 (start with this one)

The base checkpoint. Strongest LoRA + workflow ecosystem of any 2026 SDXL fine-tune.

• Architecture: SDXL 1.0 fine-tune (~2.6M images, balanced anime/cartoon/furry/pony, balanced safe/questionable/explicit)
• File size: 6.46 GB (fp16 pruned, .safetensors) on Civitai · 6.94 GB (fp16 full) on the LyliaEngine HF mirror
• VRAM: 6 GB minimum, 8–12 GB recommended at 1024px (general SDXL guidance — the author doesn't publish a Pony-specific minimum)
• Hash (AutoV2): 67AB2FD8EC · 882,604 version downloads · 71,837 reviews, "Overwhelmingly Positive"
• License: modified Fair AI Public License 1.0-SD — see the license section below
• Download from Civitai · Hugging Face mirror (full fp16) · Author's HF account

Pruned and full are the same model — pruned removes optimizer/training metadata, identical inference quality. Grab pruned for inference, full only if you plan to merge or fine-tune downstream.

2. SDXL VAE (companion file)

The decoder Pony ships with. There is no separate "Pony VAE" — community parlance for "the SDXL VAE that comes alongside Pony."

• Architecture: SDXL VAE
• File size: 335 MB (LyliaEngine HF mirror) · 319 MB (bundled on the Civitai card — same VAE, packaged differently)
• License: CreativeML Open RAIL++-M (released alongside SDXL 1.0)
• Download sdxl_vae.safetensors from Hugging Face

Skip this file if you're using a derivative checkpoint with a baked-in VAE — both Pony Realism and CyberRealistic Pony bake the VAE in.

3. Pony PDXL Negative Embeddings — zPDXL3

Drop-in quality booster that replaces the long score string. Pony-only — won't work on stock SDXL, NoobAI, Illustrious, or Flux.

• File size: 264.15 KB (.safetensors)
• Hash (AutoV2): 1727C6CDDF · 68,291 downloads · 7,159 reviews, Very Positive
• Trigger: zPDXL3 in the negative prompt (variants exist — XXX Rating forces NSFW, PG Rating forces SFW)
• Download from Civitai

Author's pitch: "no need to use score this, score that — just use the embeddings." V3 also removes Pony's built-in censoring on explicit prompts. Use both positive and negative variants together at strength 1.0–2.0 if you want both.

Pony LoRAs worth adding next

Pick one detailer, one style. Stacking two detail LoRAs from the same family halos.

4. Wonderful Details — v1.0

Sharp-edge detailer for facial fidelity, fabric texture, and micro-contrast.

• 129.56 MB · Hash: D8690A3B0D · 58,900 downloads, 224 reviews, Very Positive
• Recommended weight: 1.4 (range 1.2–2.0; >2.5 starts to halo)
• Trigger words: none
• Download from Civitai

Skip this if you're already running CyberRealistic Pony or Pony Realism — both creators bake similar detail in.

5. HIGHLY DETAILED STYLE — HDETAILED v2.0

Heavy style + detail elaboration. Pairs with other LoRAs without conflict.

• 649.69 MB · Hash: 6FFB0FBE67 · 5,000 downloads, 92 reviews, Positive
• Recommended weight: 0.3–0.9 · CFG 2.5–5
• Trigger word: aaArtemus
• Download from Civitai

Heavier hand than Wonderful Details. Pick this if you want detail you can see on the thumbnail.

6. Pony V6 Art Style — "CivitAI Most Collected Image"

Style LoRA trained on the 256 most-collected Pony images on Civitai across PG → XXX ratings.

• 217.91 MB · Hash: 8212F4A00B · 989 downloads, 12 reviews, Positive (small sample)
• Trigger: score_9 or any individual tag from the training images
• Download from Civitai

Use as a "what's hot on Civitai right now" stylistic baseline. Not a substitute for a specific artist LoRA — for that, grab the next one.

7. Artist Style | PonyXL | SDXL | Lora — v6.0

Largest artist-style LoRA pack for Pony XL by download volume.

• File size: not verified this session (Civitai card behind NSFW gate in some regions)
• Recommended weight: 0.7–1.0 with Euler a
• Download from Civitai

Re-verify download count and file size after Civitai login — the latest snapshot we could read showed ~211k downloads dated June 2024.

Variant checkpoints (use one of these instead of vanilla V6)

You don't have to start with vanilla V6. If you want a more predictable model or photoreal output, swap in one of these. They keep Pony's tag and score system, so your LoRAs and prompts still work.

8. AutismMix_pony

Tames Pony V6's prompt-instability — better hands, less style swing, no need for "schizo negatives."

• 6.46 GB (fp16 pruned) · Hash: 821AA5537F · 223,598 version downloads, 31,531 reviews, Overwhelmingly Positive
• Same score string as V6. Add source_anime, BREAK per the creator. Drop schizo negative embeds — the creator says they make output worse.
• Variants on the same page: AutismMix_pony (LoRA training), AutismMix_confetti (general use), AutismMix_DPO, AutismMix_Lightning_4step, AutismMix_Lightning_8step
• Download from Civitai

Pick this if your vanilla V6 outputs are inconsistent and you don't want to fight the model.

9. Pony Realism v2.3

Photoreal Pony — closest you'll get to RealVis-style output while keeping Pony tags and LoRA compatibility.

• 6.62 GB (fp16) · Hash (v2.2): 7C97ECF786 · 205,789 v2.2 downloads, 15,794 reviews, Overwhelmingly Positive
• Settings: DPM++ 2M SDE Karras (or Euler a / DPM2 a) · 30+ steps · CFG 6–7 · CLIP skip 2 · 1024px+
• Use female/male instead of woman/man per the creator. Avoid DPM++ 2M Karras (creator warning). Pair with Pony Realism Enhancer LoRA for max sharpness.
• VAE is baked in — you can skip the SDXL VAE file for this checkpoint
• Download from Civitai

10. CyberRealistic Pony — v18.0 CoreShift

Cinematic photoreal Pony with baked-in VAE. Latest version is paid Early Access; v17 and earlier are free.

• 6.46 GB (fp16 pruned) · Hash (v18): 1D580C1C3F · 6,800 v18 downloads, 183 v18 reviews; older free versions have many thousands more
• Settings: DPM++ SDE Karras (or DPM++ 2M Karras / Euler a) · 30+ steps · 896×1152 or 832×1216 · CFG 5 · CLIP skip 2
• ADetailer with face_yolov9c.pt for face refinement
• Download from Civitai · Hugging Face backup

If v18 is still in Early Access when you read this, drop down to v17 — it's free and 95% of the way there.

Required settings for Pony V6 (do these once)

These apply to the base V6 and every Pony-derived checkpoint. Get them wrong and the model looks broken.

• CLIP skip 2. A1111 / Forge: Settings → "show all pages" → search clip → add clip_skip_stop_at_last_layers to Quick Settings list → reload UI → set the new top-bar slider to 2. ComfyUI: drop a CLIPSetLastLayer node set to −2 between your CLIP loader and the encoder.
• Sampler / steps: Euler a at 25 steps (author default). DPM++ 2M SDE Karras at 30+ steps for photoreal output (Pony Realism / CyberRealistic recipe).
• Resolution: 1024×1024 native. Other supported: 832×1216, 1216×832, 896×1152, 1152×896. At non-1024 sizes Pony sometimes generates duplicate subjects — add solo, 1girl, or 1boy to the positive prompt.
• Prompt structure: lead with the full score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up string, then source_pony / source_furry / source_cartoon / source_anime, then your rating_* tag, then your subject. Don't add masterpiece, hd, or best quality — they aren't trained tokens, and stacking them degrades output.
• Negative prompt: the author says Pony is "designed to not need negative prompts in most cases." For NSFW, score_4, score_5, score_6 in negative tightens quality (Pony Realism / CyberRealistic recipe).

License and monetization — read before you publish images

Pony Diffusion V6 XL ships under a modified Fair AI Public License 1.0-SD. The base license text lives at freedevproject.org/faipl-1.0-sd. The Pony-specific modification, verbatim from the Civitai card:

"You are not permitted to run inference of this model on websites or applications allowing any form of monetization (paid inference, faster tiers, etc.). This applies to any derivative models or model merges."

What that means in plain English:

• Personal and research use: fine. Run it locally, generate whatever you want.
• Commercial inference on Civitai or Hugging Face: explicitly granted by the author.
• Paid generators, paid tiers, paid apps elsewhere: banned without written permission. Email contact@purplesmart.ai to request it.
• Generated images: the upstream FAIPL text says outputs aren't covered by the license, but the monetization clause restricts how those outputs reach a paying audience on a hosted service. Self-host, charge nothing, you're safe. Build a paid Pony generator without permission, you're not.
• Derivative checkpoints (AutismMix, Pony Realism, CyberRealistic Pony): downstream FAIPL restrictions apply because they're trained or merged on top of Pony V6.

The Hugging Face LyliaEngine mirror's YAML metadata says license: cdla-permissive-2.0 — that's mirror metadata, not the operative license. The Civitai card and the README prose are primary.

Where to put each file

Drop each file in the right folder, then refresh the model dropdown (A1111 / Forge) or restart the workflow (ComfyUI).

Or run LocalForge AI, where Pony V6 is preset with CLIP skip 2 already wired up and the score string scaffolded into the prompt template — no manual settings dance.

What to do next

• Want the full setup walk-through? How to run Pony Diffusion locally — the install, the first prompt, troubleshooting CLIP skip and score-string failures.
• Looking for the best Pony-family checkpoints ranked? Best Pony Diffusion models — vanilla, photoreal, anime, anthro variants compared.
• Need an NSFW-focused list? Pony Diffusion NSFW models — explicit-content checkpoints and LoRAs ranked.
• Building out a wider SDXL stack? Stable Diffusion models download list — the broader SDXL + SD 1.5 universe Pony fits inside.