How to Run Pony Diffusion Locally (Full Guide)

Why Pony Needs Its Own Guide

Pony Diffusion V6 XL is an SDXL fine-tune, so the install path is the same as any SDXL model — Forge, ComfyUI, or A1111. But the settings are not. If you skip CLIP skip 2, you'll get blobs. If you type masterpiece, best quality instead of the score_9 string, you're prompting a model that doesn't understand those words. This guide covers the Pony-specific settings that generic SD install tutorials miss.

The Steps

1. Check Your Hardware

You need an NVIDIA GPU, 8 GB+ VRAM, and ~20 GB of free disk space.

Install the latest NVIDIA drivers from nvidia.com. You also need Python 3.10.x (check "Add to PATH" during install) and Git for Windows. If any of these are missing, Forge won't launch.

2. Install Stable Diffusion WebUI Forge

Forge is the recommended UI — it's faster than A1111 on 6 GB GPUs and actively maintained.

One-click: Download the latest release from Forge on GitHub. Extract, run update.bat, then run.bat. The package includes Git and Python — no separate install needed.

Git clone: Open a terminal where you want the project, run git clone https://github.com/lllyasviel/stable-diffusion-webui-forge.git, then run webui-user.bat. First launch takes 20–30+ minutes while it downloads PyTorch and other dependencies. That's normal.

ComfyUI alternative: If you want node-based control, grab the portable build from ComfyUI's GitHub. Steeper learning curve, more power. Everything below still applies — just put files in ComfyUI's folder structure instead.

3. Download Pony V6 XL

Grab the "V6 (start with this one)" build — it's the checkpoint with 840,997 downloads and the full LoRA ecosystem.

Download from Civitai or the Hugging Face mirror. Drop the .safetensors file into your Forge models/Stable-diffusion/ folder. If colors look washed out later, add vae-ft-mse-840000-ema-pruned.safetensors to models/VAE/ — but try without it first.

After copying, click Refresh in the Forge checkpoint dropdown. Select the Pony file.

If you want Pony V6 running without assembling Python, Git, and CUDA by hand, LocalForge AI ships with the folder layout pre-wired — you still pick models and run offline.

4. Set CLIP Skip to 2

This is non-negotiable. Without CLIP skip 2, Pony V6 produces low-quality output.

In Forge: go to Settings → search for Clip skip → set to 2 → click Apply settings. In ComfyUI, use a CLIP loader node with stop_at_clip_layer: -2. The model author states this explicitly on the Civitai card — skip this and you'll wonder why everything looks wrong.

5. Write Your First Prompt

Use the full score string — score_9 alone is much weaker than the full chain.

Here's the template:

score_9, score_8_up, score_7_up, score_6_up, source_anime, rating_safe, 1girl, smile, blue eyes, park background

• Quality tags: score_9, score_8_up, score_7_up, score_6_up (add score_5_up, score_4_up if you want)
• Source tag: source_anime, source_pony, source_furry, or source_cartoon — picks the style
• Rating tag: rating_safe for SFW, rating_explicit for NSFW, rating_questionable for middle ground
• Subject tags: Describe what you want after the system tags

Set Euler a, 25 steps, CFG 7, and a resolution like 1024×1024 or 832×1216. Hit Generate.

Don't use masterpiece, best quality, or HD — those are for non-Pony models. Don't paste negative prompt walls from Discord — Pony V6 is designed to work without them.

6. Fix Common Problems

Most Pony issues come from three mistakes: wrong CLIP skip, wrong prompt tags, or not enough VRAM.

• OOM (out of memory) on 6 GB: Add --medvram to your launch args in webui-user.bat. Costs ~20–30% speed but cuts VRAM usage by 40–50%. Also try lowering resolution to 832×832.
• OOM on 8 GB with hi-res fix: Drop hi-res fix first. Add --xformers to launch args.
• Blurry/blobby output: Check CLIP skip — it must be 2.
• Tags don't seem to work: Make sure you're using the full score_9, score_8_up, … string, not just score_9.
• Model doesn't show up: Wrong folder. It belongs in models/Stable-diffusion/, not models/checkpoints/ or anywhere else. Hit Refresh.
• Washed-out colors: Try adding a VAE file to models/VAE/ and selecting it in the UI.

7. What About Pony V7?

V7 is a different model family (AuraFlow, not SDXL). Don't mix V6 settings with V7.

Pony V7 base exists on Civitai and Hugging Face with GGUF quantized weights for lower VRAM. But the tooling is newer, quality tags are weaker, and V7.1 is still planned. Start with V6, prove your workflow, then experiment with V7 as a separate install. See Best Pony Diffusion Models for the full V6 vs V7 comparison.

Quick Comparison

What to Do Next

• Pick NSFW models — Pony Diffusion NSFW models covers rating_explicit workflows, NSFW LoRAs, and which merges work best for explicit content.
• Compare all Pony models — Best Pony Diffusion models ranks V6, V7, AutismMix, and Illustrious-lane picks with download stats.
• Broaden to all SD NSFW — Best NSFW Stable Diffusion models if you want SD 1.5 photoreal or non-Pony SDXL options.
• Try a managed stack — LocalForge AI handles the install plumbing so you start generating instead of debugging Python paths.