Load ImageReward Verifier

CLIP can tell you whether an image matches its prompt, but it's got no taste. ImageReward is the verifier that does: it's a reward model trained on human preference rankings of text-to-image outputs, so its score reflects "would a person actually like this?" rather than just "does it match the words."

That's the middle ground in this pack's three-judge lineup, and for most workflows it's the sweet spot. It's not as cheap as the CLIP verifier - but it's a rounding error next to the Qwen VLM's VRAM and speed costs, and its judgment is meaningfully closer to human taste.

How it works

Load ImageReward Verifier loads ImageReward-v1.0, the reward model from the 2023 paper "ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation" (THUDM). Internally it's a BLIP model plus a learned scoring head that was trained to rank images the way human raters did. Give it an image and a prompt, and it returns a preference score. First run auto-downloads the weights from HuggingFace (roughly 2GB) and caches them, so you don't need to source anything by hand.

One quirk worth knowing: the node wraps the .to(device) call in a try/except because some ImageReward models resist being moved directly - if that silently fails, the model just stays where it loaded and scoring still works. Not something you need to fix, just something that explains odd behavior if you're digging through logs.

Inputs

• ir_verifier_id - exactly one choice, ImageReward-v1.0. It's fixed; there's no model zoo to browse here.
• device - cuda or cpu, default cuda. Leave it alone.

Output

One wire: image_reward_verifier_instance (type IR_VERIFIER), which feeds Inference Time Scaler's loaded_image_reward_verifier input.

Install

Pack install, same as the other verifier loaders - ComfyUI Manager search "ComfyUI-InferenceTimeScaling", or:

cd ComfyUI/custom_nodes
git clone https://github.com/YRIKKA/ComfyUI-InferenceTimeScaling
cd ComfyUI-InferenceTimeScaling
pip install -e .

Honest expectations

Two things to know before you trust its rankings blindly. First, its scores aren't on a pretty 0–1 scale - the shipped Flux example shows ImageReward handing out ~1.8 for decent images, and bad ones can land negative. Judge by relative ranking, not by reading the number. Second, its training data is overwhelmingly SD 1.5-era generated images, so its sense of quality was calibrated on older aesthetics. On SDXL or Flux it still works and generally agrees with people, but don't treat it as gospel for styles that didn't exist when it was trained.

For most setups this is the verifier you'll actually reach for: sharper taste than CLIP, a fraction of the cost of the Qwen VLM. Run it with the CLIP verifier for a solid two-judge ensemble and only add the VLM when you're chasing a hero image and can afford the wait.