| --- |
| license: apache-2.0 |
| task_categories: |
| - text-to-image |
| tags: |
| - synthetic |
| - qwen-image |
| - latent-to-pixel |
| - l2p |
| size_categories: |
| - 10K<n<100K |
| --- |
| |
| # L2P-Clean — curated (prompt, image) pairs for Latent-to-Pixel transfer on Qwen-Image-2512 |
|
|
| Aesthetic-curated synthetic dataset for the L2P distillation recipe |
| ([arXiv:2605.12013](https://arxiv.org/abs/2605.12013)): the source diffusion model's |
| own outputs, cleaned "to the bones" so the pixel decoder fits an already-smooth, |
| high-quality teacher manifold. |
|
|
| ## Provenance |
|
|
| Merged and de-conflicted from three independent synthetic runs (each its own prompt |
| corpus — prompt ids are namespaced with a `raw_` / `part0_` / `dataset_` prefix because |
| the numeric ids collide across runs but refer to different prompts): |
|
|
| | source | prompts | seeds/prompt | raw images | |
| |--------|--------:|-------------:|-----------:| |
| | `shauray/l2p-raw` | 6,172 | 3 | 18,516 | |
| | `shauray/l2p-part0` | 4,421 | 1 | 4,421 | |
| | `shauray/l2p-dataset` | 2,100 | 2 | 4,200 | |
|
|
| (one truncated source shard skipped). All prompt **texts** are unique across the three runs. |
|
|
| ## Curation pipeline (`clean_l2p.py`) |
| |
| Following Krea's "quality over quantity" philosophy — a **human-preference** model |
| (PickScore), deliberately **not** a LAION-aesthetic predictor (which biases toward |
| soft/symmetric/blurry "AI look"): |
| |
| 1. **PickScore** every image against its own prompt (CLIP-ViT-H backbone). |
| 2. **Per-prompt seed selection**: keep the top-2 seeds per prompt above a global |
| PickScore floor (p12 — permissive, to preserve the teacher manifold for distillation, |
| cutting only clear failures). Matches the paper's 2-seeds-per-prompt recipe. |
| 3. **pHash near-duplicate removal** across survivors (Hamming ≤ 6), catching identical |
| look-alikes across prompts/seeds. |
| 4. **Diversity audit**: k-means over CLIP embeddings → inverse-frequency |
| `sampling_weights.jsonl` so over-represented visual modes are down-weighted in the |
| dataloader (fixes repetition without deleting anything). |
|
|
| ## Format |
|
|
| WebDataset shards (`shard-*.tar`), each sample is `<key>.png` + `<key>.txt` (the prompt). |
| Images are ~1472×1104. Plus: |
| - `manifest.jsonl` — `{id, prompt, files}` per sample |
| - `sampling_weights.jsonl` — `{id, cluster, weight}` inverse-frequency sampling weights |
|
|
| ## Final size |
|
|
| - **18,505 curated (prompt, image) pairs** across 75 shards |
| - per-source survivors: {'raw': 11449, 'part0': 3648, 'dataset': 3408} |
|
|
