Load T5 Model 🐼

What it is

The loader half of this pack's T5-based prompt expander. The README name-drops it directly: support for "T5 models such as roborovski/superprompt-v1" - and right now that's the only model this node actually offers. SuperPrompt is a small T5 model fine-tuned specifically to take a short, plain prompt and expand it into a longer, more descriptive one - the kind of thing you'd otherwise do by hand or with a much bigger LLM. This node loads it; T5 Text to Prompt is the node that actually runs generation against it.

How it works

Straightforward loader: pick the model (there's only one choice right now), pick a device, and optionally wire in a quantization config if you need to shrink the model's memory footprint. It hands back a T5_MODEL object that T5 Text to Prompt consumes.

Inputs and outputs

• model - a single-choice dropdown, roborovski/superprompt-v1. Not much of a decision to make here yet, but the field is structured as an enum, which suggests room for more T5 models to be added later.
• device - cuda (default) or cpu.
• quantizationConfig - optional, QUANTIZATIONCONFIG type, fed from the T5 Quantization Config node. Leave this unconnected and the model loads at full precision; wire in a config if you're squeezed on VRAM and want to run it in 8-bit or 4-bit instead. See that node's article before you touch this - the bitsandbytes fields it exposes aren't beginner-friendly, and SuperPrompt is small enough that most people won't need to quantize it at all.

Output: T5_MODEL, wired into T5 Text to Prompt.

Installing it

Part of the same pack as every other node here:

cd ComfyUI/custom_nodes
git clone https://github.com/zhongpei/Comfyui_image2prompt

Or install via ComfyUI Manager, searching "Comfyui_image2prompt." (The README's own clone command is typo'd to a hyphenated URL that 404s - use the underscored one above.) Restart ComfyUI afterward.

The README doesn't call out superprompt-v1 in its explicit model-download list the way it does for the vision-language models, but the same general pattern applies to every model this pack uses: it should pull down automatically from Hugging Face the first time you run the node, and if that doesn't happen, huggingface-cli download roborovski/superprompt-v1 --local-dir <path> gets you there manually, with HF_ENDPOINT=https://hf-mirror.com available if you're downloading from behind the Great Firewall.

Common issues

This is a small model - quantization is usually unnecessary. SuperPrompt is a lightweight T5, nowhere near the size of the 7-8B vision-language models elsewhere in this pack. Unless you're running on genuinely constrained hardware, load it at default precision and skip the quantization config entirely; it adds complexity (bitsandbytes-specific settings) for a problem you probably don't have.

Mixing this up with the pack's other prompt-generation path. This pack ships two separate text-generation systems: this T5 loader plus T5 Text to Prompt, and the Qwen-based Loader Text to Prompt Model plus Text to Prompt. They're not interchangeable - a T5_MODEL only wires into T5-family nodes, and a TEXT2PROMPT_MODEL only wires into the Qwen-family ones. If a connection won't drag between two nodes, that's ComfyUI telling you the types don't match, not a bug.

CPU fallback works but is slow for anything beyond quick tests - device: cuda is the default for a reason, even though the model itself is small relative to the rest of the pack.