💎 Pony Ultimate Pro

Pony Diffusion V6 is one of the most capable models you can run, and one of the most annoying to prompt. Every generation needs that exact score-tag string - score_9, score_8_up, score_7_up in the positive, score_5_up, score_4_up in the negative - because a training quirk made the whole list effectively mandatory, not a nice-to-have. Forget a tag and your output quietly gets worse.

PonyUltimatePro is a single node that does all that for you and then some. It's the all-in-one generation node from the SeppoUK/PonyUltimate pack: it loads your checkpoint, encodes your prompt, runs the sampler, decodes, and - its genuinely useful trick - runs a face detail pass on the result, all inside one box. If you're tired of wiring the same seven-node chain every time you test a Pony merge, this collapses it.

What's actually happening under the hood

Read the source and it's a standard ComfyUI chain compressed into one gen() call. It loads the checkpoint with Comfy's own load_checkpoint_guess_config, prepends the score tags to both prompts, appends your chosen style's keywords (pick "Photographic" and you get raw photo, realistic skin texture, bokeh, 35mm lens appended), builds an empty latent at your chosen aspect ratio, and calls the same internal KSampler a normal workflow uses. The name doesn't call an API and needs no key - "ultimate" just means everything's in one place.

The face fix is the part worth caring about. It pulls in the Impact Pack's FaceDetailer plus an Ultralytics YOLO detector (face_yolov8m.pt) to find faces, crop and resample them at proper resolution, then composite them back - the detect-crop-refine loop that's the ComfyUI answer to ADetailer. Pony's "small faces come out mangled" problem is exactly what that loop exists to solve, and the node hardwires it in so you don't have to configure FaceDetailer's two dozen inputs yourself.

The inputs that matter

• ckpt_name - pick any Pony V6-family checkpoint you have installed. The default points at Pony\realismByStableYogi_v60FP16.safetensors, a realism merge you almost certainly don't own (more on that below).
• positive / negative - your normal tag prompt. Do not type the score tags; they get prepended for you. The default negative (source_furry, source_pony) already steers away from the furry/pony aesthetic, which is the right instinct for most V6 uses.
• style - a dropdown of keyword presets: Cinematic, Photographic, Digital Art, Anime, Vintage, and a few more. "None" leaves your prompt untouched.
• aspect_ratio - SDXL-safe presets (832x1216 portrait, 1344x768 landscape, etc.) so you're not generating off-native sizes.
• face_detailer + face_denoise - on/off, and how hard the detail pass alters the face. The 0.55 default is a sane starting point; nudge it up only if faces still look soft.
• seed, steps, cfg, sampler_name, scheduler - the usual knobs, defaulting to dpmpp_2m_sde / karras / CFG 5.5 / 30 steps, which is a reasonable Pony baseline.

It outputs IMAGE - the final image, or a side-by-side before/after if show_comparison is Yes - plus MODEL and VAE pass-throughs so you can keep sampling that same checkpoint elsewhere.

Installing it

In ComfyUI Manager, search for PonyUltimate and install, or clone directly:

cd ComfyUI/custom_nodes
git clone https://github.com/SeppoUK/PonyUltimate

Then restart ComfyUI. The pack itself ships zero Python dependencies - there's no requirements.txt - so the real install is its hidden dependency: the Impact Pack (ComfyUI-Impact-Pack). And here's the catch the README glosses over: since Impact Pack v8.0, the UltralyticsDetectorProvider class this node calls lives in the separately-installed ComfyUI-Impact-Subpack. Install both, or the face fix won't exist at all. You'll also need the face_yolov8m.pt detection model available to the Subpack - it's the same Bingsu/adetailer YOLO weight the A1111 world uses.

Where people get burned

The default checkpoint doesn't exist on your machine. The node's dropdown is hardcoded to a StableYogi realism merge you have to go download, and if you hit Generate anyway, it throws "Checkpoint not found" and hands you a black 64x64 image. Pick a Pony checkpoint you actually have from the dropdown first.

The face fix fails silently. If the Impact Subpack or face_yolov8m.pt is missing, the node logs a warning in the console and quietly returns the base image. You'd never know the detail pass didn't run unless you watch the terminal.

And the "Pony" in the name is a constraint, not a theme. Those score tags do nothing useful on Illustrious, Anima, or plain SDXL - this node's whole point is automating a V6-era workflow. If you've already moved off the Pony family, skip it. If you're still exploring Pony V6 and its endless realism merges, it's a legit set-and-forget way to do it.