📥 Load checkpoint Pony (+Download from CivitAi)

If you generate on Pony, you already know CivitAI is where the checkpoints live - AutismMix, 2DN-Pony, PonyRealism, the whole family of V6 merges. This node saves you the round trip. You pick a Pony checkpoint from a curated in-node list (about 74 of them), and if the file isn't on your disk yet, it downloads it from CivitAI and loads it. Same behavior as a normal Load Checkpoint node, just with CivitAI as the source instead of your local folder.

Pony is CivitAI's own base-model category for a reason - SDXL LoRAs and tooling only half-work on it because the training diverged so far - so having a Pony-specific picker that grabs the right files is more useful than it sounds.

How it works

On run, the node checks whether the checkpoint you selected exists locally; if not, it pulls it from CivitAI into ComfyUI/models/checkpoints/Bjornulf_civitAI/pony/, then loads it like any checkpoint. First run on a new model pays the download cost, and after that it's a normal local file you can also load with a plain Load Checkpoint node.

The inputs and outputs that matter

• image - the model picker, shown as a thumbnail gallery of Pony checkpoints (roughly 74 entries). This is the input you choose.
• civitai_token - your CivitAI API token, from your account settings. Needed only to authorize the download; blank is fine once the file is local.

Outputs are model, clip, and vae for your KSampler / CLIP Text Encode / VAE Decode, plus name and civitai_url strings for metadata.

Actually getting good output from Pony

The node loads the model; it can't prompt it for you, and Pony is famously particular. Two things that trip up newcomers, worth knowing before you blame the checkpoint:

• The score tags are mandatory, and you use the whole string. Prefix your positive prompt with score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up. This looks redundant, but a training quirk baked the sequence in as the quality signal - individual score tags don't do the job. This is Pony-specific; it does nothing on SDXL base, Illustrious, or FLUX.
• Prompt in booru tags, not sentences. The heavy e621/Danbooru tagging largely overwrote SDXL's plain-English understanding, so 1girl, blue_hair, from_behind, forest beats a flowing description. Source tags like source_anime (positive) or source_furry, source_pony (negative) steer the aesthetic.

For style consistency, most people reach for a V6 merge like AutismMix rather than base Pony - vanilla Pony can hand you four different styles from one prompt.

How to install it

ComfyUI Manager: search Bjornulf_custom_nodes, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/justUmen/Bjornulf_custom_nodes
pip install -r Bjornulf_custom_nodes/requirements.txt

Restart ComfyUI. The civitai-py dependency from that requirements file is what powers the download, so don't skip it. Windows portable users: .\python.exe -m pip install ....

Common issues

The download needs a valid token, and gated or early-access CivitAI models can still refuse. If a first run hangs, check the token before anything else. Pony checkpoints are SDXL-sized (6.5GB+), so expect a real pause on the initial download of each new model - it's not frozen.

Also worth a heads-up: a lot of Pony content on CivitAI is NSFW, and much of the model catalog reflects that. And don't reach for Pony V7 out of habit - it moved to a different architecture, landed poorly, and has almost no ecosystem; V6 XL and its merges are still the practical choice. This node is Pony-specific; use the SDXL, SD1.5, or FLUX siblings for those bases.