PaCo-RL: Advancing Reinforcement Learning for Consistent Image Generation with Pairwise Reward Modeling
Paper
•
2512.04784
•
Published
•
23
Data and Model collection for PaCo-RL
Note Pairwise images with visual consistency annotation
Note Benchmark to evaluate human preference for visual consistency
Note Specialized MLLM for human preference alignment on visual consistency
Note Lora adapter of PaCo-Reward-7B. Merge it with Qwen2.5-VL-7B-Instruct to obtain PaCo-Reward-7B
Note Lora adapter for FLUX.1-dev's transformer
Note Lora adapter for FLUX.1-Kontext-dev's transformer
Note Lora adapter for Qwen-Image-Edit's transformer