I also seperately provide just the prompts in prompts.json
keys are the image_id, and the values are the captions generated
Captions generated by moondream: vikhyatk/moondream2
Latents generated by SDXL VAE: madebyollin/sdxl-vae-fp16-fix
Embeddings generated by SigLIP: hf-hub:timm/ViT-SO400M-14-SigLIP-384
Original dataset: common-canvas/commoncatalog-cc-by
Latents f32 and embeddings are f16 bytes
Compute cost: 16x3090 for 3 day. Approximately.
