Skip to content
Bud's Wiki

liuzhuang1234 data findings — long captions + prompt rewrite, equal weighting, epochs over volume, narrow hi-res data

TL;DR

Section titled “TL;DR”

A short empirical thread from Zhuang Liu (FAIR) summarizing four data-pipeline findings from a recent image-generation training run: (1) train on long captions and rewrite short user prompts at inference, rather than training on short captions; (2) equal weighting across curated datasets is a strong default; (3) once the mix is diverse, trading off image count for more epochs barely costs quality; (4) strong high-resolution generation does not require broad coverage in the high-res slice. No paper, no model release — just four claims with no quantitative anchors attached.

Key claims

Section titled “Key claims”
  • Training on long captions plus inference-time prompt rewriting beats training on short captions, even when the inference prompts themselves are short [tweet].
  • Equal weighting is a strong default for mixing curated datasets [tweet].
  • Given a diverse-enough dataset mix, fewer unique images with more epochs does not hurt performance much [tweet].
  • Strong high-resolution performance does not require broad coverage in the high-resolution training data [tweet].

Method

Section titled “Method”

None disclosed. The tweet is a bulleted list of findings; no model name, scale, ablation tables, or evaluation suite is attached. The accompanying image is described as a media attachment with no extra detail in the fetched payload.

Results

Section titled “Results”

No numbers. Each claim is reported as a qualitative direction (“strong default”, “doesn’t hurt much”).

Why it’s interesting

Section titled “Why it’s interesting”

The long-caption-train + short-prompt-rewrite-at-inference pattern matches what Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer does explicitly — Z-Image generates five caption tiers per image (long/medium/short/tags/simulated-user) and conditions on world-knowledge meta-info for the long captions — and what Generating an Image From 1,000 Words: Enhancing Text-to-Image With Structured Captions (FIBO) pushes to the extreme by training only on ~1,000-word structured JSON captions. Liu’s tweet is the same recipe stated as folklore. The hi-res-doesn’t-need-broad-coverage claim is the more contrarian one and connects to long-running questions in Diffusion training efficiency about where compute should go in a multi-resolution curriculum.

See also

Section titled “See also”