MR-IQA-2 couples a multimodal Actor, a frozen FLUX.2-klein-4B Editor, and a frozen E5 Judge. Masked credit assigns reasoning and rating rewards only to their eligible completion tokens.

MR-IQA-2 masked-credit training overview

Actor-3 — Better consistency

Actor-3 is the consistency-oriented variant trained with pairwise source/edited-image consistency, retained edit utility, and the preservation-first v9 prompt. The checkpoint, exact training/test prompts, and a directly runnable single-image script are published in actor-3/.

bash

huggingface-cli download RobinY99/MR-IQA-2 \
  --include 'actor-3/**' \
  --local-dir mr-iqa-2-actor-3
python mr-iqa-2-actor-3/actor-3/inference.py /absolute/path/to/input.jpg \
  --model mr-iqa-2-actor-3/actor-3 \
  --local-files-only

Quick start: one image, one GPU

Download the runnable code bundle, create the two pinned inference environments, and provide one image:

bash

python -m pip install 'huggingface-hub==0.34.4'
huggingface-cli download RobinY99/MR-IQA-2 \
  --include 'code/**' \
  --local-dir mr-iqa-2-hf
cd mr-iqa-2-hf/code
bash scripts/setup_envs.sh --profile inference
python examples/quick_start.py /absolute/path/to/input.jpg --gpu 0

Actor, Editor, and Judge run sequentially in separate processes on the selected GPU. The Actor solution is forwarded verbatim to the Editor; the Judge reports J0, J1, and J1-J0. No HTTP service is started.

On hosts with multiple CUDA toolkits, select the toolkit used for runtime extension compilation:

bash

python examples/quick_start.py /absolute/path/to/input.jpg \
  --gpu 0 \
  --cuda-home /usr/local/cuda

Outputs are written to outputs/quick_start/: actor_raw.txt, assessment.json, edited.png, evaluation.json, and result.json. The release path was smoke-tested end to end on one NVIDIA A6000 (48 GB), with J0=3.42, J1=4.12, and J1-J0=+0.70.

PLCC/SRCC performance

Actor-only rating performance on six generalization datasets. Each entry is PLCC / SRCC; Average is the unweighted macro mean.

Model	KonIQ-10K	SPAQ	LIVE-W	AGIQA-3K	KADID-10K	CSIQ	Average
MR-IQA	0.949 / 0.931	0.892 / 0.897	0.899 / 0.883	0.804 / 0.732	0.672 / 0.683	0.767 / 0.732	0.831 / 0.810
MR-IQA-2	0.937 / 0.917	0.900 / 0.899	0.893 / 0.863	0.809 / 0.739	0.667 / 0.669	0.824 / 0.785	0.838 / 0.812

MR-IQA values are from the released Qwen3-VL-2B result in the MR-IQA paper. MR-IQA-2 uses the released masked-credit E5 Actor at step 1,455. Exact coefficients and valid-row counts are in the checkpoint results.

Released folders

actor-3/ (Better consistency): five-epoch Qwen3.5-4B Actor with pairwise consistency, retained edit utility, and the low-level v9 prompt;
actor-2(continued training from judge model)/ (Recommended): masked-credit Actor with component KL, trained for one Actor epoch;
actor/: masked-credit E5 Actor, step 1,455;
judge/: frozen E5 Judge, step 725;
editor/: FLUX.2-klein-4B;
code/: runnable single-image inference bundle.

actor was initialized from Qwen3.5-4B and trained for five epochs. actor-2 was initialized from a pretrained rating-only Judge trained for five epochs, then trained as an Actor for one epoch with masked credit and component KL.

Actor and Judge load with AutoModelForImageTextToText.from_pretrained using the corresponding subfolder. The Editor loads from editor/ with Flux2KleinPipeline.from_pretrained.

A real Actor → Editor sample, including the exact completion and provenance, is available in examples/actor_editor/sample_0001.json. Training, evaluation, and deployment code is available on RobinY99/MR-IQA-2.

Citation

bibtex

@misc{li2026mriqa2,
  title={MR-IQA-2: Faithful Image Quality Reflection via Fine-Grained Credit Assignment},
  author={Yuan Li and Youyuan Lin and Chenhui Chu and Shin'ya Nishida},
  year={2026},
  eprint={2608.18579},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2608.18579}
}

RobinY99/MR-IQA-2

MR-IQA-2 logoMR-IQA-2

Actor-3 — Better consistency

Quick start: one image, one GPU

PLCC/SRCC performance

Released folders

Citation

RobinY99/
MR-IQA-2

MR-IQA-2 logo
MR-IQA-2