PyTorch implementation of Multimodal LLM-Guided Semantic Correction in Text-to-Image Diffusion
Run the following script to start inference:
bash run.sh| Argument | Description |
|---|---|
--save_dir |
Directory to save the generated results |
--device |
Device to use for inference (cuda, cpu, etc.) |
--num_images_per_prompt |
Number of images to generate per prompt |
--num_inference_steps |
Total denoising steps; higher means better quality but slower generation |
--model_path |
Path to the diffusion model (e.g., HunyuanDiT) |
--vl_model_path |
Path to the vision-language model for evaluation (e.g., Qwen2.5-VL) |
--seed |
Random seed for reproducibility |
--process_steps |
Number of self-reflection optimization steps |
--process_steps_interval |
Interval between each self-reflection step during generation |
--process_start |
Fraction of inference steps after which to start self-reflection (e.g., 0.1) |
--process_end |
Fraction of inference steps to stop self-reflection (e.g., 0.9) |
--use_self_reflection |
Whether to enable self-reflection during image generation |
Generated outputs are saved under the specified --save_dir directory. Example structure:
results/
├── image_0.png
├── image_1.png
├── ...
└── prompts.txt
Make sure the following libraries are installed:
diffuserstransformerstorch
Or install from requirements.txt