You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

UniReFuse: MSRS full-COT SFT epoch4

This repository contains the full model weights used to initialize ImageFusion-R1 GRPO/RL.

Protocol:

degraded infrared image input;
degraded visible image input;
infrared restoration CoT and clean infrared image;
visible restoration CoT and clean visible image;
fusion CoT and clean fused image.

Training summary:

base architecture: Lumina-mGPT-2.0-Omni;
dataset: balanced MSRS Level1/Level2 SFT set, 6,000 items;
checkpoint: epoch4, corresponding to five completed training epochs;
objective: autoregressive CE over text, image tokens, and image structure tokens;
intended downstream use: initialization for the repository's MSRS full-protocol GRPO trainer.

The repository intentionally excludes optimizer states. It is suitable for fresh RL initialization, but not for resuming the original SFT optimizer. The original Lumina tokenizer/base assets, MoVQGAN checkpoint, ImageFusion-R1 code, Qwen reward model, and MSRS data are distributed separately.

Verify downloaded files with sha256sum -c SHA256SUMS.

Downloads last month: -

Safetensors

Model size

8B params

Tensor type

BF16