Multi-view Perspective Image Generation from BEV with Test-time Controllability and Generalizability (ACM MM24, Poster)

This work aims to address the multi-view perspective RGB generation from text prompts given Bird-Eye-View(BEV) semantics. Unlike prior methods that neglect layout consistency, lack the ability to handle detailed text prompts, or are incapable of generalizing to unseen view points, MVPbev simultaneously generates cross-view consistent images of different perspective views with a two-stage design, allowing object-level control and novel view generation at test-time. Specifically, MVPbev firstly projects given BEV semantics to perspective view with camera parameters, empowering the model to generalize to unseen view points. Then we introduce a multi-view attention module where special initialization and denoising processes are introduced to explicitly enforce local consistency among overlapping views w.r.t. cross-view homography. Last but not the least, MVPbev further allows test-time instance-level controllability by refining a pre-trained text-to-image diffusion model. Our extensive experiments on NuScenes demonstrate that our method is capable of generating high-resolution photorealistic images from text descriptions with thousands of training samples, surpassing the state-of-the-art methods under various evaluation metrics. We further demonstrate the advances of our method in terms of generalizability and controllability with the help of novel evaluation metrics and comprehensive human analysis. Our code and model will be made available.

• Use pip install -r requirements.txt to install all dependencies. Notably, make sure torch.__version__ >= 2.1.0.

You should build a dataset for this project based on NuScenes dataset, or you can download datasets that we've already created. If you make it right, the dataset should be organized as following structure which is pretty simple:

.
└── DATASET-PATH/
    ├── train.csv
    ├── valid.csv
    └── data/
        ├── [TOKEN].pt (≈10MB)
        └── ...

Here are some guidance to create this dataset on your own:

• Get to target folder cd ./create_dataset
• Optional & Not recommend: Compile Cython code with sh ./scripts/cybfs_setup.sh. You can extract some extra data wiht this enabled, but that's not really used in our final version.
• Configure your dataset through config.py.
• Run it by python build_dataset.py which typically takes over 20 hours (by default).

• First, cd ./scripts/py_scripts, we provide several python scripts here, and you can have all jobs done here.
• Make sure those pretrained models (SD, Controlnet, etc.) in HuggingFace are available for you. Optional: You can download all models we need to local (MVPbev/weights/pretrained) by python download_pretrained.py. You can configure which way you prefer (load pretrained model from local or remote) in MVPbev/configs/*.py
• Step 1. We finetune pretrained Stable Diffusion to be able to generate stree-view images. Start finetuning by python train.py --s 0 --n [EXP_NAME], the training program will update some test results and checkpoints saved in MVPbev/logs/[EXP_NAME](auto created).
• Step 2. Train Multi-view Attention Module by python trian.py --s 1 --n [EXP_NAME].
• Please note we finetune SD only while leave Controlnet untouched.
• You can always modify concrete training setups in config file MVPbev/finetune_SD.py and MVPbev/train_MVA.py.

• Configure path to your model checkpoint to be tested in config file MVPbev/configs/test_config.py.
• cd ./scripts/py_scripts and python test.py --n [EXP_NAME], this will test your model in valid set (you can set how many samples you want to use in config file).
• All test results will be saved to MVPbev/test_output (auto created).
• You can look through generated samples in MVPbev/test_output/test_results/sample_*.pt, which is a dict.
• Load and browse generated results with torch.load().

• Clone CVT code to local and organize it as following structure. You need a pretrained CVT weight (you can directly download the CVT checkpoint we trained) and download labels: cvt_labels_nuscenes_v2.

.
├── cross_view_transformers/
│   ├── ...
│   └── model.ckpt
├── cvt_labels_nuscenes_v2/
│   ├── scene-*
│   └── ...
└── MVPbev/
    └── ...

• If your files are in same structure as above, you can simply evaluate your model by python evalute.py --n [EXP_NAME], which will compute all metrics used in our paper, or you need pass two extra params: python evaluate.py --n [EXP_NAME] --cvt_code_path [PATH_TO_CVT_CODE] --cvt_ckpt_path [PATH_TO_CVT_CKPT].
• All generated samples in MVPbev/test_outputs/[EXP_NAME] will be evaluated, and you will see something like this:

• We demonstrated how to use our training-free instance control via a notebook in MVPbev/multi_objs_color_control, including usages in detail and implementation of all our evaluation metrics.
• You can also see a csv file in that dir, that's a list of test samples we used in our paper. ⚠️ But, please note not all data in that list exist in dataset we provided now. Becauese, in this new version, we expanded our dataset by set min_gap_distance=8 in config file to sample more data. The old version of dataset that contains all those samples is not compatible to this version of code. In this way, we can not directly share that to you. So, if you want to test on all those samples, please re-build dataset with setting min_gap_distance=10, or you can generate those specific samples only since you've already known their tokens. It's a smarter way but requires coding for a little bit, or you can simply set LOAD_MISSING=True in the notebook, you'll need full NuScenes dataset in your device in that case.

• MVPbev Checkpoints
• Pretrained CVT model (for evaluation)

• You can directly use our created dataset (≈15 GB zipped, contains ≈10,000 Sets × 4MB ≈ 40GB).

We originally implemented this project basing on following repos:

• MVDiffusion: Stable Diffusion training & inference, CAA implementation.
• Diffusers: Stable Diffusion + Controlnet forward code.

[TO-BE-UPDATED]