Pony Pull

What Pony Pull Actually Is

Pony Pull is a tool and workflow people use with Stable Diffusion models, specifically the Pony Diffusion family. It's essentially a pipeline helper that automates model loading, prompt parsing, and batch image generation from command-line or script interfaces. The Pony models in question are the fine-tuned SDXL-based models (v6 and earlier) that have a very distinct art style, and Pony Pull makes it easier to work with them without going through a GUI every time. I set up Pony Pull about a year ago after spending too much time manually tweaking parameters in WebUI. The basic flow is: you point it at your model file, feed it a text file full of prompts (one per line), and it churns out images at whatever resolution and sampler settings you configure. It handles batching, which is the main draw. Instead of running one prompt at a time through a browser interface, you let it run overnight. The configuration lives in a YAML or JSON file. You specify your model path, output directory, steps, CFG scale, sampler type, and resolution presets. I usually keep mine at 64 steps with DPM++ 2M Karras and a CFG of 4.5, which is about right for Pony v6 models. Lower CFG than you'd use with standard SDXL because Pony models are trained to respond well at lower guidance values.

Installation and Setup

You can find Pony Pull on GitHub, typically under repositories related to Pony Diffusion tooling. Clone it into your extensions or scripts folder, install the Python dependencies listed in requirements.txt, and make sure your CUDA or MPS setup is working before you start. It depends on diffusers and sometimes on ComfyUI nodes depending on which fork you're using, so check which variant you downloaded. The ComfyUI-native version tends to be more stable for batch work. I ran into a specific issue where Pony Pull would crash mid-batch on prompt 47 every single time. Turned out one of my prompt files had a hidden zero-width character in it from copying text out of a web form. The tokenizer would choke on it. My workaround was running my prompt list through a simple Python script first that strips non-printable characters, then passing the cleaned file to Pony Pull. Added about 30 seconds to my workflow but stopped the crashes entirely.

Pitfalls and What No One Tells You

The biggest thing people miss is that Pony Pull doesn't do any prompt upscaling or resolution correction for you. If you paste a prompt meant for 1024x1024 into a 832x1216 batch run, you're going to get distorted compositions. The Pony models are particularly sensitive to aspect ratio mismatches because their training data is heavily curated around specific resolutions. I learned this after wasting three hours on a batch that looked weirdly stretched. Now I validate my aspect ratios against the model's training specs before running anything. Another thing: VRAM management. Pony Pull loads the model into memory once and reuses it across all prompts in a batch, which is efficient. But if you're also running something else on the same GPU, you will OOM. I keep mine isolated on a dedicated card now. Also, the default timeout settings are generous but not infinite. Long batches on slow prompts can hit the timeout and drop out silently. Check your logs. The tool has real limitations. It doesn't handle negative prompts elegantly in all versions, batch ordering isn't always deterministic across runs, and if your model file has any custom LoRA embedded in its metadata, Pony Pull may or may not respect it depending on the version. The most recent forks have improved on the LoRA handling, but I've seen inconsistent results between the diffusers-based and ComfyUI-based implementations. If you need reliable LoRA application, stick to the ComfyUI variant.

For people who need more control over individual prompts within a batch, Pony Pull isn't the right tool. You'd be better off with a workflow that lets you tweak parameters per-prompt rather than applying one global config across the whole batch. Pony Pull shines when you have 50 to 200 variations of similar prompts and just want to generate them all consistently without sitting at a screen.