JourneyDBCache: Precomputed Latents for Boomer Text-to-Image Training

A partial cache of 3,955,000 processed samples containing DC-AE image latents and Gemma text hidden states derived from JourneyDB-style image-caption pairs. Created by akrao9, this dataset is intended to reduce frozen encoder costs during Boomer model training. The cache generation was halted prematurely, skipping 849 source samples.

Use Cases

• Train text-to-image models based on precomputed image latents and text embeddings.
• Fine-tune diffusion models based on cached feature representations to save compute resources.
• Benchmark training efficiency improvements based on using pre-extracted latent features.
• Experiment with multimodal latent spaces based on combined image and text hidden states.

Strengths

• Contains 3,955,000 precomputed samples, providing a substantial base for training.
• Specifically designed to remove frozen encoder computational cost from the Boomer training pipeline.
• Caches both image latents (DC-AE) and text hidden states (Gemma), offering a multimodal feature set.

Limitations

• Description metadata is limited; actual data quality and structure require manual inspection after download.
• Row count for the full intended cache is unknown, as the run was partial.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

huggingface

Collection Method

Precomputed from JourneyDB-style image-caption WebDataset shards.

Freshness

Last updated 2026-05-23 05:16:58; freshness should be verified.