Sovereign 8s.is mirror of standardgalactic/milr
  • Python 84.4%
  • HTML 9.1%
  • Cuda 3.8%
  • Shell 2%
  • C++ 0.4%
  • Other 0.2%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-02-12 11:44:48 +08:00
fig init push 2025-10-16 11:34:21 +08:00
src add details 2025-10-17 20:15:20 +08:00
.gitignore init push 2025-10-16 11:34:21 +08:00
README.md Update author order in citation 2026-02-12 11:44:48 +08:00
requirements.txt add gpt4o NVILA eval 2025-08-04 21:40:00 +08:00

MILR: Improving Multimodal Image Generation via Test-Time Latent Reasoning ⚡️ICLR 2026

teaser map

We introduce MILR, a test-time method that performs joint reasoning over text and image in a unified latent space. It searches over vector representations of discrete text/image tokens using policy gradients guided by an image-quality critic, instantiated within the MUG framework (which supports language reasoning before image synthesis). MILR optimizes intermediate model outputs as the latent space, requiring no fine-tuning and operating entirely at inference time.

Installation

conda create -n latentseek python=3.10
conda activate latentseek
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

#install Geneval configs
#You may meet package counters, it doesn't matter
pip install -U openmim
mim install mmengine mmcv-full==1.7.2
cd src/geneval
./evaluation/download_models.sh "<OBJECT_DETECTOR_FOLDER>/"
git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection; git checkout 2.x
pip install -v -e .

cd ../rewards
./evaluation/download_models.sh "<OBJECT_DETECTOR_FOLDER>/"

If you meet error with flash_attn==2.7.2.post1, you can refer to the https://github.com/Dao-AILab/flash-attention/releases to download.

Usage

We support different kinds of reward types. And we test on three benchmarks: Geneval, T2I-CompBench, Wise

Geneval

cd src
bash scripts/geneval_both.sh

The bash file

#!/bin/bash
PATH_TO_DATA="prompts/geneval/evaluation_metadata.jsonl"
PATH_TO_MODEL="deepseek-ai/Janus-Pro-7B"
output_dir="./geneval_results/long_results" #self create the dir
optimize_mode="both"  # or "image"
reward_model_type="geneval"
text_k=0.1 
image_k=0.01 
lr=0.01
max_text_steps=30
max_image_steps=30
max_both_steps=30

# === set log file name ===
if [ "$optimize_mode" = "text" ]; then
    LOG_FILE="$output_dir/${optimize_mode}_tk${text_k}_lr${lr}_ts${max_text_steps}.txt"
elif [ "$optimize_mode" = "image" ]; then
    LOG_FILE="$output_dir/${optimize_mode}_ik${image_k}_lr${lr}_is${max_image_steps}.txt"
else
    LOG_FILE="$output_dir/${optimize_mode}_tk${text_k}_ik${image_k}_lr${lr}_bs${max_both_steps}.txt"
fi

# === train script ===
CUDA_VISIBLE_DEVICES=1 python main_janus.py \
    --dataset "$PATH_TO_DATA" \
    --model_name_or_path "$PATH_TO_MODEL" \
    --output_dir "$output_dir" \
    --optimize_mode "$optimize_mode" \
    --reward_model_type "$reward_model_type" \
    --lr "$lr" \
    --text_k "$text_k" \
    --image_k "$image_k" \
    --max_text_steps "$max_text_steps" \
    --max_image_steps "$max_image_steps" \
    --max_both_steps "$max_both_steps" \
    --device "cuda" \
    > "$LOG_FILE" 2>&1 &
  • optimize_mode: The mode of optimization, you can choose from both, image or text.
  • reward_model_type: the reward model used for optimize, you can check in the main_janus.py file
  • text_k: the ratio of text tokens for optimization
  • image_k: the ratio of image tokens for optimization
  • lr: the learning rate
  • max_text_steps: the steps of text optimization
  • max_image_steps: the steps of image optimization
  • max_both_steps: the steps of both optimization

Different Rewards

For SelfReward, you can run the script:

bash scripts/geneval_self_reward.sh 

For UnifiedReward, you can run the script:

bash scripts/geneval_unified_reward.sh 

remeber you download the CodeGoat24/UnifiedReward-qwen-7b model.

For MixedReward, you can run the following script:

cd rewards/MixedReward
git clone https://github.com/IDEA-Research/GroundingDINO.git ## follow the guide of https://github.com/IDEA-Research/GroundingDINO
mkdir reward_weights
wget -q https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha/groundingdino_swint_ogc.pth
huggingface-cli download microsoft/git-large-vqav2 --repo-type model --local-dir git-large-vqav2
cd ../..
bash scripts/unified_reward_geneval.sh 

For GPT-4o reward, you should first set the api key in main_janus.py, and run the following script:

bash scripts/geneval_gpt4o_reward.sh 

T2I-CompBench

For T2I-CompBench, we use this benchmark from https://github.com/Karine-Huang/T2I-CompBench And you should follow the guidence of the repo to install the environment, and put the weights into the scripts/T2ICompBench

then you can run the scripts in scripts/T2I-CompBench as follows:

bash scripts/T2I-CompBench/MetricReward_T2ICompBench_color.sh 

WISE

For WISE, we use this benchmark from https://github.com/PKU-YuanGroup/WISE. You also should install the environment following the guidences and fill the right api key into rewards/wise_reward.py.

Then you can run the script as follows:

bash scripts/Wise/wise_reward_wise_cultural_common_sense.sh 

Citation

@article{mi2025milr,
  title={MILR: Improving Multimodal Image Generation via Test-Time Latent Reasoning},
  author={Mi, Yapeng and Zhao, Yanpeng and Li, Hengli and Li, Chenxi and Wu, Huimin and Ma, Xiaojian and Zhu, Song-Chun and Wu, Ying Nian and Li, Qing},
  journal={arXiv preprint arXiv:2509.22761},
  year={2025}
}

Contact