proxima_cena_grounded_original_dataset

Paired next-scene dataset for reference-conditioned image generation training.

Layout

Four sources, each with the same structure:

dsN_<name>/
  images_A/   # control (reference) image
  images_B/   # target image + <same-stem>.txt caption

Pairs are matched by identical file stem across images_A/images_B.

subset	pairs	note
ds1_recortados	2,900	random subset (seed 42)
ds2_poxima_v2	5,400	main subset (90% of source, seed 42)
ds3_comikontext	2,900	random subset (seed 42)
ds4_contexto_curado	1,255	manually curated pairs (trained with 3x repeats)

Total: 12,455 pairs.

Captions

Each caption is a rich standalone description of the target image (B) only — framing first, then subjects with appearance and precise pose, action and contact, environment and lighting. No trigger words, no editing instructions, no style/medium/quality words, no references to the control image. Generated with the grounded-scene preset of vidcap (single-image mode): Gemini 3.1 Flash Lite in batch mode, with refused items filled by Qwen3-VL-30B-A3B / Gemma 4 31B / Seed 2.0 Lite via OpenRouter.

The intent: what the caption specifies is the target state; everything it does not mention (identity, style, palette, world details) is meant to be inherited from the reference image by the model being trained.