minDALL-E on Conceptual Captions

minDALL-E, named after minGPT, is a 1.3B text-to-image generation model trained on 14 million image-text pairs for non-commercial purposes.

Environment Setup

• Basic setup PyTorch == 1.8.0 CUDA >= 10.1
• Other packages pip install -r requirements.txt

Model Checkpoint

• Model structure (two-stage autoregressive model)
  • Stage1: Unlike the original DALL-E [1], we replace Discrete VAE with VQGAN [2] to generate high-quality samples effectively. We slightly fine-tune vqganimagenetf16_16384, provided by the official VQGAN repository, on FFHQ [3] as well as ImageNet.
  • Stage2: We train our 1.3B transformer from scratch on 14 million image-text pairs from CC3M [4] and CC12M [5]. For the more detailed model spec, please see configs/dalle-1.3B.yaml.
• You can download the pretrained models including the tokenizer from this link. This will require about 5GB space.

Sampling

• Given a text prompt, the code snippet below generates candidate images and re-ranks them using OpenAI's CLIP [6].
• This has been tested under a single V100 of 32GB memory. In the case of using GPUs with limited memory, please lower down numcandidates to avoid OOM. ```python from matplotlib import pyplot as plt import clip from dalle.models import Dalle from dalle.utils.utils import setseed, clip_score

device = 'cuda:0' set_seed(0)

prompt = "A painting of a monkey with sunglasses in the frame" model = Dalle.from_pretrained('minDALL-E/1.3B') # This will automatically download the pretrained model. model.to(device=device)

Sampling

images = model.sampling(prompt=prompt, topk=256, # It is recommended that topk is set lower than 256. topp=None, softmaxtemperature=1.0, num_candidates=96, device=device).cpu().numpy() images = np.transpose(images, (0, 2, 3, 1))

CLIP Re-ranking

modelclip, preprocessclip = clip.load("ViT-B/32", device=device) modelclip.to(device=device) rank = clipscore(prompt=prompt, images=images, modelclip=modelclip, preprocessclip=preprocessclip, device=device)

Plot images

images = images[rank] plt.imshow(images[0]) plt.show() ``` - If you want to use a complete python code for sampling, please see examples/sampling_ex.py - If you want to play with an interactive demo, please see examples/samplinginteractivedemo.ipynb. Before using this, you may need to install ipywidgets.

Samples (Top-K=256, Temperature=1.0)

• "a painting of a {cat, dog} with sunglasses in the frame"

  

  

• "a large {pink, black} elephant walking on the beach"

  

  

• "Eiffel tower on a {desert, mountain}"

  

  

Quantitative Results

• We have validated minDALL-E on the CC3M validation set (in-distribution evaluation) and MS-COCO (zero-shot evaluation).
• For CC3M, we measure the cosine similarity between image and text representations from the pretrained CLIP model (ViT-B/32), referred to as CLIP-score.
• For MS-COCO, we compute FID between 30K generated and real samples from MS-COCO 2017, where we randomly choose 30K captions from COCO as in DALL-E. We select the best out of 32 candidates by CLIP re-ranking.

| Model | CC3M:CLIP-score (higher is better) | MS-COCO:FID-30K (lower is better) | |:------|----:|----:| |VQGAN [2] | 0.20 | - | |ImageBART [7]| 0.23 | - | |DALL-E [1] | - | 27.5 | |minDALL-E | 0.26 | 14.7 |

Transfer Learning Examples

• minDALL-E, which is pre-trained on noisy text supervisions, could be transferable to class-conditional and unconditional generation tasks. To validate this, we simply fine-tune it on ImageNet over 8 epochs in the case of class-conditional generation and unconditional generation.
• The commands below fine-tune the pretrained DALL-E. It takes about 36 hours on 8 V100 GPUs. ``` # unconditinoal image generation for imagenet (256x256) python examples/transferlearningex.py -d=configs/transfer-imagenet-uncond-gen.yaml -u=[MODELCKPT] -r=[RESULTPATH] --n-gpus=[NUM_GPUS]

class-conditinoal image generation for imagenet (256x256)

python examples/transferlearningex.py -d=configs/transfer-imagenet-clscond-gen.yaml -u=[MODELCKPT] -r=[RESULTPATH] --n-gpus=[NUM_GPUS] `` * We compute FID-50K between 50K generated samples and all ImageNet training samples, where we use top-k=256 and softmax temperature=1.0 for generation. All results are obtained without the rejection sampling. Interestingly, our model achieves very competitive performance with baselines, even thoughminDALL-E` is fine-tuned in a few epochs.

| Model | Params | FID-50K(class-cond.) | FID-50K(uncond.) | |:-----|----:|----:|----:| |VQ-GAN | 1.4B | 15.78 | - | |ImageBART | 3.5B | 21.19 | - | |minDALL-E | 1.3B | 15.55 | 37.58 |

BibTex

If you find this repository useful in your research, please cite: @misc{kakaobrain2021minDALL-E, title = {minDALL-E on Conceptual Captions}, author = {Saehoon Kim, Sanghun Cho, Chiheon Kim, Doyup Lee, and Woonhyuk Baek}, year = {2021}, howpublished = {\url{https://github.com/kakaobrain/minDALL-E}}, }

References

• [1] Ramesh et al. Zero-Shot Text-to-Image Generation, ICML 2021.
• [2] Esser et al. Taming Transformers for High-Resolution Image Synthesis, CVPR 2021.
• [3] Karras et al. A Style-Based Generator Architecture for Generative Adversarial Networks, CVPR 2019.
• [4] Sharma et al. Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning, ACL 2018.
• [5] Changpinyo et al. Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts, CVPR 2021.
• [6] Radford et al. Learning Transferable Visual Models From Natural Language Supervision, ICML 2021.
• [7] Esser et al. ImageBART: Bidirectional Context with Multinomial Diffusion for Autoregressive Image Synthesis, NeurIPS 2021.
• [8] https://github.com/karpathy/minGPT

Licenses

• The source codes are licensed under Apache 2.0 License.
• The stage2 pretrained weights are licensed under CC-BY-NC-SA 4.0 License.

Contact

We hope that minDALL-E helps various projects in research-oriented institutes and startups. If you would like to collaborate with us or share a feedback, please e-mail to us, contact@kakaobrain.com

Limitations

Although minDALL-E is trained on a small set (14M image-text pairs), this might be vulnerable to malicious attacks from the prompt engineering to generate socially unacceptable images. If you obersve these images, please report the "prompt" and "generated images" to us.