README.md

metadata

library_name: imscore
datasets:
  - carolineec/CyclePrefDB-I2T
tags:
  - model_hub_mixin
  - pytorch_model_hub_mixin

port of cyclereward

There are three variants of the CycleReward model:

1. CycleReward-Combo (This model)
2. CycleReward-I2T
3. CycleReward-T2I

This model has been pushed to the Hub using the PytorchModelHubMixin integration:

• Code: https://github.com/Abhinay1997/imscore, cyclereward branch
• Paper: 2506.02095
• Docs: [More Information Needed]

To use this model:

• install imscore from my fork

pip install git+https://github.com/Abhinay1997/imscore@cyclereward

• use the model as follows

from imscore.cyclereward.model import CycleReward
available_models = ["NagaSaiAbhinay/CycleReward-Combo" "NagaSaiAbhinay/CycleReward-T2I", "NagaSaiAbhinay/CycleReward-I2T"]
model_id = available_models[0]
model = CycleReward.from_pretrained(model_id)

prompts = "a photo of a cat"
pixels = Image.open("cat.jpg")
pixels = np.array(pixels)
pixels = rearrange(torch.tensor(pixels), "h w c -> 1 c h w") / 255.0

# prompts and pixels should have the same batch dimension
# pixels should be in the range [0, 1]
# score == logits
score = model.score(pixels, prompts) # full differentiable reward)