Load Image Reward Score Model 🐼

What it is

The loader half of the pack's image-scoring pair. On its own it doesn't do much - it exists purely to hand THUDM/ImageReward to the Image Reward Score node, which does the actual scoring and culling work. If you're setting up an automated "generate several, keep the best" step in your graph, this is the first node you place, not the last.

ImageReward itself is a real, still-relevant tool - trained on human preference comparisons rather than a simple aesthetic classifier, and reward-model scoring of exactly this shape (rank a folder of generations, keep the good ones) is still something people build tools around today, years after ImageReward first shipped.

How it works

It's about as simple as a loader gets: pick a device, it loads the ImageReward model weights, and hands back an IMAGEREWARD_MODEL object.

Inputs and outputs

• device - cuda (default) or cpu. That's the entire input surface.

Output: IMAGEREWARD_MODEL, which only wires into one place - the Image Reward Score node's model input.

Worth noting: this node is flagged as an output node in its schema, even though functionally it's a loader. That doesn't change how you use it - you still wire its output forward into Image Reward Score - but it means ComfyUI will treat it as a valid graph endpoint on its own if you ever need one, which isn't typical for a loader node.

Installing it

Ships as part of the full Comfyui_image2prompt pack:

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

Or search "Comfyui_image2prompt" in ComfyUI Manager. (The README's own clone command has a typo - a hyphenated URL that 404s - use the underscored one above.) Restart ComfyUI after installing.

The README doesn't put ImageReward in its explicit Hugging Face model-download list the way it does moondream and the deepseek-vl models, but the same general behavior should apply: the model downloads automatically the first time you run this node, and if that fails, you can fetch it manually - the README's mirror advice (HF_ENDPOINT=https://hf-mirror.com) applies here too if you're behind the Great Firewall.

Common issues

This node alone does nothing visible. If you load it and run your graph without wiring the output into Image Reward Score, you won't see an error, but you also won't get anything useful - it's a pure dependency for that other node.

First run downloads weights, same as everything else in this pack. Don't be alarmed if the first execution takes a while; check your console log to confirm it's downloading rather than stuck.

cuda vs. cpu. Scoring a large batch of images on CPU will work but will be noticeably slower than the generation step that produced them - if you're building an automated cull-the-batch pipeline and it feels sluggish, check this is actually set to cuda and that your GPU has the headroom for it alongside whatever else is loaded.