---
configs:
- config_name: synth_chars
  data_files:
  - split: train
    path: "data/synth_chars/*.parquet"
- config_name: big_liminal
  data_files:
  - split: train
    path: "data/big_liminal/*.parquet"
- config_name: deepfashion
  data_files:
  - split: train
    path: "data/deepfashion/*.parquet"
- config_name: ffhq
  data_files:
  - split: train
    path: "data/ffhq/*.parquet"
- config_name: flux_assorted_bulk
  data_files:
  - split: train
    path: "data/flux_assorted_bulk/*.parquet"
- config_name: flux_assorted_bulk_2
  data_files:
  - split: train
    path: "data/flux_assorted_bulk_2/*.parquet"
- config_name: imagenet_synthetic
  data_files:
  - split: train
    path: "data/imagenet_synthetic/*.parquet"
- config_name: imdb
  data_files:
  - split: train
    path: "data/imdb/*.parquet"
- config_name: mannequins_v10
  data_files:
  - split: train
    path: "data/mannequins_v10/*.parquet"
- config_name: mannequins_v7
  data_files:
  - split: train
    path: "data/mannequins_v7/*.parquet"
- config_name: full
  data_files:
  - split: train
    path: "data/*/*.parquet"
---

# diffusion-pretrain-set-ft1-1024

1024px (2x) upscale of [AbstractPhil/diffusion-pretrain-set-ft1](https://huggingface.co/datasets/AbstractPhil/diffusion-pretrain-set-ft1).

# WARNING

MUCH OF THIS DATA WAS MODEL UPSCALED USING RAPID UPSCALERS.

THIS IS NOT CONSISTENTLY HIGH FIDELITY NOR IS IT EVEN CLOSE TO FAIR FIDELITY AT TIMES.

PLEASE use this ONLY for pretraining, new concepts, and simple design purposes ONLY. HEAVILY PRUNE FOR FINETUNING.

Thank you, good luck my friends.

# Details

- Model: realesr-general-x4v3 (SRVGG Compact, spandrel), fp16, native 4x forward
  -> bicubic+antialias retarget to 2x (supersampled).
- Selected by speed/fidelity Pareto on real data: rw_lpips 0.021,
  LR-consistency PSNR 41.4 (hallucination guard).
- Re-encoding: webp q95.
- All Image columns upscaled; every other column passed through unchanged.
- Images with min edge >= 1024px kept verbatim (original bytes/encoding).
- All image columns normalized to typed Image() features (several source subsets stored untyped raw bytes).
- Shards: 2048 source rows each, row groups of 512.