DataSet Source for Training EVA-VLIP + SDXL as Visual Autoencoder (LAION-COCO Dataset)

[shanpoyang654]

Description:

I'm working on training EVA-VLIP + SDXL as a visual autoencoder using the LAION-COCO dataset from Hugging Face. I've noticed that:
The dataset contains parquet files with image URLs and captions, not the actual images

Many URLs appear to be broken or inaccessible

This results in only a small fraction of images being successfully downloaded

Questions:
For your training runs, did you use:

Only the successfully downloaded images from LAION-COCO's URLs?

Or did you have access to a more complete/pre-downloaded version of the dataset?
If using an alternative dataset source:

Could you recommend where to obtain a more reliable version of LAION-COCO with higher URL availability?

Are there other suitable datasets you'd recommend for training SDXL conditioned on visual embeddings?
For the image-text pairs that fail to download:

Did you implement any fallback strategies (e.g., using cached versions from other sources)?

Or did you simply exclude these samples from training?

[CnFaker]

I’d also like to know how many GPUs were used to train EVA-VLIP + SDXL, and how long the training took.