PhillMagine320: A Text-to-Image Training Dataset for Diffusion Model Adapters

A 320-row dataset created by author ayjays132, last updated on 2026-05-21. It is a small, curated text-to-image training dataset combining the original PhillMagine corpus with Gemini-synthesized samples. The dataset was used to train a universal text-conditioning adapter for diffusion models.

Use Cases

• Training a text-conditioning adapter for diffusion models based on the described image-prompt pairs.
• Fine-tuning image generation models using a combined corpus of original and synthetic samples.
• Benchmarking text-to-image generation performance on a small, curated dataset.
• Studying the effect of augmenting training data with AI-synthesized samples on model conditioning.

Strengths

• Dataset size is explicitly stated as 320 total rows, with a defined split of 288 for training and 32 for testing.
• Data sources are documented, specifying 120 rows from 'phillmagine120' and 200 rows from 'gemini_synth_v1'.
• The dataset has a clear, specialized purpose for training a universal text-conditioning adapter.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but the total data scale is small at 320 rows, which may limit suitability for some training tasks.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Combination of an original corpus ('phillmagine120') and AI-synthesized samples ('gemini_synth_v1').

Freshness

Last updated 2026-05-21 09:34:31; freshness should be verified.