README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - reward-model
  - poster
  - graphic-design
  - image-quality-assessment
  - preference-learning
  - qwen3-vl
pipeline_tag: image-to-text
library_name: transformers

PosterReward: Unlocking Accurate Evaluation for High-Quality Graphic Design Generation

Overview

PosterReward is a dedicated reward modeling framework for poster assessment. It builds a 70k poster preference dataset from multi-MLLM consensus and introduces specialized models for poster quality evaluation across five dimensions:

1. Foundational Visual Quality
2. AI Artifacts
3. Textual Accuracy
4. Prompt Fidelity
5. Aesthetic Value

Available Models

This repository contains three model variants, all built on Qwen3-VL-8B:

Model	Path	Type	Description
PosterReward Analyser	PosterReward_analyser/	Generative VLM	Generates detailed multi-dimensional analysis of poster images
PosterReward Scorer	PosterReward_scorer/	Scalar Reward Model	Takes analysis + image and produces a scalar reward score
PosterReward-Lite	PosterReward-Lite/	Scalar Reward Model	Simplified pointwise scorer that omits the analysis module for faster inference

PosterReward (Full Pipeline)

The full PosterReward pipeline is a two-stage analysis -> scoring process:

1. PosterReward Analyser generates a detailed textual analysis across five quality dimensions.
2. PosterReward Scorer takes the analysis together with the image and outputs a scalar reward score.

PosterReward-Lite

A simplified variant that directly predicts a scalar reward from the image and prompt, without requiring a separate analysis step. Faster inference at the cost of slightly lower accuracy.

Benchmark Data

This repository also hosts the PosterRewardBench benchmark images:

File	Description	Size
PRB_basic_images.tar.gz	PosterRewardBench-Basic images (1,034 images from Flux, Flux-Krea, SD3.5-L)	~1.1 GB
PRB_advanced_images.tar.gz	PosterRewardBench-Advanced images (2,446 images from Seedream-3.0, Seedream-4.0, Qwen-Image-Lightning)	~736 MB

Download and extract these archives into the poster_reward_bench/ directory of the code repository.

Quick Start

Environment Setup

git clone https://github.com/MeiGen-AI/PosterReward.git
cd PosterReward

cd swift && pip install -e . && cd ..
pip install msgspec "qwen_vl_utils>=0.0.14" torchvision diffusers pillow

PosterReward-Lite (Fast Pointwise Scoring)

from swift.llm import PtEngine, InferRequest

model_path = "path/to/PosterReward-Lite"  # or download from this repo
engine = PtEngine(model_path, max_batch_size=64, task_type='seq_cls', num_labels=1)

messages = [
    {"role": "user", "content": "<image>Your poster description prompt here."},
    {"role": "assistant", "content": ""}
]
request = InferRequest(messages=messages, images=["path/to/poster.png"])

resp_list = engine.infer([request])
score = resp_list[0].choices[0].message.content
print(f"Reward Score: {score}")

Full PosterReward (Two-Stage Pipeline)

# Edit model paths in inference_posterreward.sh, then:
bash inference_posterreward.sh

Results

Pointwise Reward Models on PosterRewardBench

Model	MMRB2 ↑	HPDv3 ↑	PRB-Basic ↑	PRB-Ad ↑
ImageReward	53.0	58.6	60.7	49.3
PickScore	57.6	65.6	66.7	44.1
HPSv2	55.0	65.3	70.8	43.7
HPSv3	58.5	76.9	72.9	41.2
PosterReward-Lite	60.5	77.1	83.9	85.0
PosterReward	59.6	77.8	86.7	86.0

Citation

Coming Soon!

Acknowledgments

• Thanks to our collaborators and affiliated institutions.
• Thanks to the open-source community and prior reward modeling research.