| --- |
| license: other |
| tags: |
| - text-to-image |
| - multi-subject |
| - layout-control |
| - diffusion |
| --- |
| |
| # PRISM dataset — Stage 1 images |
|
|
| Transfer mirror of the PRISM dataset's Stage-1 generated images: 200,000 |
| PNG images generated by FLUX.1-dev from Stage-0 scene designs. |
|
|
| **This repo is a staging area.** It currently holds the image half of the |
| dataset only; the record index (entries) is not uploaded yet. |
|
|
| ## Layout |
|
|
| | Path | Contents | |
| |---|---| |
| | `prism/stage1_images/shard_0000.tar` … `shard_0099.tar` | 100 tar shards, ~2.7 GB each (267 GB total) | |
| | `prism/stage1_images/manifest.jsonl` | one line per member: which shard it is in, its original path, and its size | |
|
|
| Each tar member keeps the original path of the file, so a shard can be |
| extracted in place: |
|
|
| ``` |
| tar -xf shard_0000.tar -C <dest> # restores prism/stage1_flux_images/... as it was |
| ``` |
|
|
| To find where one image lives: |
|
|
| ```bash |
| grep '"000123_02.png"' manifest.jsonl |
| ``` |
|
|
| ### What is inside the shards |
|
|
| | | | |
| |---|---| |
| | PNG images | **200,000** (the dataset itself) | |
| | `prism/stage1_flux_images/README.md` | how Stage 1 was generated | |
| | `prism/stage1_flux_images/metadata.jsonl` | a partial earlier metadata file | |
| | `.../variant_0{0..3}/metadata.jsonl` | per-variant generation metadata — **partial** (see below) | |
| | `prism/stage1_flux_images/filtered/scenes_*.jsonl` | the Stage-0 scene designs, i.e. the prompts the images were generated from | |
|
|
| ## Why shards and not 200,000 loose files |
|
|
| The Hub allows only 128 write requests per hour per repository. The |
| preupload endpoint caps a batch at 1,000 paths, so uploading the images |
| individually needs at least 200 preupload calls plus one content-batch call |
| per 256 files plus one commit per batch — several hours of rate-limit |
| waiting. The Hub itself recommends "fewer, larger files" for datasets this |
| size. |
|
|
| ## The images |
|
|
| | | | |
| |---|---| |
| | Resolution | 1024 × 1024, RGB, PNG | |
| | Generator | FLUX.1-dev (the run's own metadata records `variant: "gguf"`) | |
| | Sampling | 20–25 steps, guidance 3.5, `max_sequence_length` 256 | |
| | Naming | `<seed_idx>_<variant_idx>.png`, e.g. `000000_00.png` | |
| | Seed range | `seed_idx` 0 – 73,999, four variants per seed | |
|
|
| Each image was generated from a Stage-0 scene design that specifies the |
| subjects, their attributes, positions and sizes; those designs are the |
| `filtered/scenes_*.jsonl` files inside the shards. |
|
|
| ## Provenance |
|
|
| | | | |
| |---|---| |
| | Source tree | `<cluster>/project_outputs/PRISM/stage1_flux_images/` | |
| | Upstream | PRISM pipeline, Stage 1 (`project_codes/PRISM/src/stage1_flux_v2.py`) | |
|
|
| ## The record index is not here |
|
|
| The complete 200,000-entry index lives in the Stage-3 output |
| (`stage3_entries_repaired/variant_0{0..3}/entries.jsonl`, 4.7 GB) and is |
| **not uploaded here**. Note that those entries carry the earlier schema and |
| the earlier Florence-2 grounding, which a full audit found unfit for |
| training. |
|
|
| The per-variant `metadata.jsonl` files are **partial** (`variant_00` and |
| `variant_02` hold 16,667 and 16,666 lines against 50,000 images each) and |
| must not be used as an index. |
|
|
