p-image-lora

PrunaTextToImageLoRA (shown as p-image-lora) is the LoRA version of the pack's text-to-image node: same hosted API, same paid-per-image model, but with a lora_weights field that lets the model adopt a custom style. The catch, and it's a big one, is what "LoRA" means here.

This is not your local LoRA ecosystem. You can't point this at a .safetensors file you trained or downloaded from CivitAI - there's no file path, no model folder, no local disk involved at all. The node takes a HuggingFace URL of the form huggingface.co/<username>/<repo>[/<filename>], and the README is explicit that the LoRA must have been trained with Pruna's own p-image-trainer. If you don't have a Pruna-trained LoRA sitting on HuggingFace, the field is dead weight. That's the honest framing: this node is for people inside Pruna's training loop, not for the wider community's LoRA library.

What you set

• prompt - same plain-English multiline input as p-image.
• lora_weights - the HuggingFace URL above. This is the field that matters, and the default is literally huggingface.co/username/repo as a reminder.
• lora_scale - influence of the LoRA, -1.0 to 3.0 in steps of 0.05, default 0.5. Start at the default; push toward 1.0+ if the style is too faint, and remember negative values push away from the style.
• aspect_ratio - same seven options as the base node (1:1 default).
• api_key - your Pruna key, or blank to use PRUNA_API_KEY.
• hf_api_token (optional) - a HuggingFace token for private repos. Only sent to the API if you actually fill it in.

Output is a single image socket, type IMAGE, ready for any Preview or Save node.

How it works

Same machinery as PrunaTextToImage - one POST to https://api.pruna.ai/v1/predictions, poll until done, download the image as a tensor. The only difference is the payload carries lora_weights and lora_scale alongside the prompt, so Pruna's server (not your machine) merges the adapter into the model before generating.

Install

ComfyUI Manager → search ComfyUI Pruna API → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/PrunaAI/comfyui-pruna-api
pip install -r comfyui-pruna-api/requirements.txt

Restart, and the nodes appear under the pruna ai category. Dependencies are just requests and Pillow - nothing heavy, no model downloads. You still need a key from the Pruna developer portal.

Where people get burned

• Assuming any LoRA works. The single most common misunderstanding. Pruna-trained + HuggingFace URL, or it silently fails / ignores the adapter.
• The placeholder default. Run it without touching lora_weights and you've asked for huggingface.co/username/repo, which won't resolve. Fill it in or don't use this node.
• Same paid-per-image economics as p-image - plus you're billed for generations that don't do what you wanted while you tune lora_scale.

It's a narrow node for a narrow audience. If you're using Pruna's trainer, it's exactly what you want. If you're not, the base PrunaTextToImage is the one that's actually useful to you.