Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 916861333 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

i1-dataset (Asuka-FM Curated Open-Domain 625K Corpus)

A balanced, high-resolution 625,000-sample open-domain text-to-image dataset built for low-compute Diffusion Transformer (DiT) pretraining on consumer hardware. This dataset was used to train the open-domain foundation models in Asuka-FM and is derived from the multi-source data curation and synthetic multi-captioning recipe introduced in i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models.


Dataset Summary & Motivation

Following empirical findings from i1 (Zeng et al., 2026), strong text-to-image generation does not require scraping hundreds of millions of unfiltered web images:

  1. Equal Dataset Weighting (i1 Finding 8): Combining diverse real, synthetic, and text-rich sources in equal proportions prevents large homogeneous corpora from dominating gradient updates.
  2. Multi-Caption Regularization (i1 Finding 9): Pairing fewer unique images with 5 distinct long synthetic captions per image provides continuous semantic regularization across epochs while cutting storage footprints to consumer NVMe/RAM budgets.
  3. Zero-Disk AVIF + Parquet Packaging: Images are normalized, aspect-ratio validated, encoded as high-quality AVIF binary payloads (quality=80), and packed into Snappy-compressed Parquet shards with schema compatibility for asuka-fm and toy-diffusion.

Dataset Composition (625,000 Samples)

The dataset contains 125,000 samples from each of five curated sources (625,000 total), filtered strictly so that $\min(\text{width}, \text{height}) \ge 512\text{ px}$:

Source Dataset Upstream Hub Repo Samples Domain Focus
FLUX-Reason-6M LucasFang/FLUX-Reason-6M 125,000 Complex spatial composition & reasoning (synthetic)
GPT-Edit-Simpler UCSC-VLAA/gpt-edit-simpler 125,000 Instruction-aligned output images (synthetic/edited)
Midjourney v6 Recap Photoroom/midjourney-v6-recap 125,000 Artistic lighting, textures & aesthetics (synthetic)
Pexels animetimm/pexels-tagger-v0-w640-ws-full 125,000 High-resolution natural photography & human anatomy (real)
TextAtlas5M CSU-JPG/TextAtlas5M 125,000 Scene text, slides & dense OCR (TextScenesHQ, TextVisionBlend, PPT2Structured)

Synthetic Multi-Captioning (zlab-princeton/i1-captions)

Every image is aligned by source key to 5 descriptive paragraph captions generated by Qwen3-VL-30B-A3B (FP8) from the official zlab-princeton/i1-captions release (captioned on center-cropped squares resized to at most $512 \times 512$).

(Note: A high-resolution subset of 396,635 samples from i1 with $\min(H, W) \ge 1024\text{ px}$ was additionally used for Stage 3 real-data Pixel DiT fine-tuning in Asuka-FM.)


Curation & Preprocessing Pipeline

Generated with hf-dataset-toolkit (HFDataPipeline) via streaming ingestion:

  1. Sliding-Window Source Balancing: Streams active datasets concurrently (max_active_sources), interleaving equal-sized slices into shuffled 10,000-sample Parquet shards (data_shard_XXXXX.parquet).
  2. Resolution & Aspect-Ratio Filtering (OnlineBucketManager):
    • Rejects images with $\min(W, H) < 512\text{ px}$.
    • Assigns each image to the nearest NovelAI aspect-ratio bucket (dimensions divisible by 64, VAE factor 8) and prunes extreme aspect ratios exceeding max_log_ar_diff = 0.35.
  3. EXIF Normalization & AVIF Compression: Applies ImageOps.exif_transpose, converts to RGB, and encodes into in-memory AVIF bytes (quality=80), cutting storage/network overhead by $>70%$ relative to WebP/PNG without degrading VAE latent reconstructions.
  4. Sequence Length Tiering: Pre-computes token length tiers ([77, 152, 227]) for batch packing, reducing transformer sequence padding overhead by up to 66%.
  5. Manifest Export: Publishes a metadata.json manifest summarizing total samples, $\ge 1024\text{ px}$ sample counts, resolution tier counts, and the full bucket_info registry.

Dataset Schema

Each Parquet shard adheres to the following PyArrow schema (retaining column names such as booru_id and tag_weight for drop-in compatibility with asuka-fm dataloaders):

Column Type Description
booru_id string Unique identifier formatted as "{source}_{key}" (e.g., fluxreason_10492).
image binary EXIF-transposed RGB image encoded as AVIF bytes (quality=80).
prompt list<string> List of 5 synthetic captions from zlab-princeton/i1-captions.
bucket_idx int32 Assigned NovelAI bucket index in metadata.json.
target_width int32 Assigned bucket width (divisible by 64).
target_height int32 Assigned bucket height (divisible by 64).
original_width int32 Raw source image width ($\ge 512$).
original_height int32 Raw source image height ($\ge 512$).
aspect_ratio float32 Raw aspect ratio (original_width / original_height).
tier int32 Sequence length tier (77, 152, or 227) measured on prompt[0].
aesthetic_tier int32 Resolution tier flag: 3 if $\min(W, H) \ge 1024\text{ px}$, 2 if $\ge 512\text{ px}$.
tag_weight float32 Per-sample loss scaling weight (uniform 1.0 for balanced corpus).

Usage Example

import io
import random
from datasets import load_dataset
from PIL import Image

# Stream directly from Hugging Face Hub without full local download
dataset = load_dataset(
    "aipracticecafe/i1-dataset",
    split="train",
    streaming=True,
)

for sample in dataset:
    # Decode in-memory AVIF bytes to RGB PIL Image
    image = Image.open(io.BytesIO(sample["image"])).convert("RGB")

    # Sample one of the 5 Qwen3-VL-30B-A3B captions per epoch (i1 recipe)
    prompt = random.choice(sample["prompt"])

    # Metadata for aspect-ratio bucketing or shifted square cropping
    orig_w = sample["original_width"]
    orig_h = sample["original_height"]
    tier = sample["tier"]
    is_1mp = sample["aesthetic_tier"] == 3

Known Characteristics & Limitations

  • aesthetic_tier Reflects Resolution, Not Aesthetics: Downstream pipelines should treat aesthetic_tier == 3 strictly as a high-resolution filter ($\min(W, H) \ge 1024\text{ px}$), not as a visual quality score.
  • Square-Native Synthetic Bias: Two of the five sources (FLUX-Reason-6M and Midjourney-v6-Recap) consist primarily of $1024 \times 1024$ square generations. When training continuous-RoPE models with aggressive shifted square cropping, off-center focal crops on square images can reduce virtual camera pan/zoom stability compared to multi-aspect illustration corpora (see Asuka-FM Section 6.7).
  • Caption Alignment & Token Tiering:
    • Synthetic captions in i1-captions were produced on center-cropped squares ($\le 512^2$); periphery objects in extreme rectangular Pexels or TextAtlas5M images may be omitted from captions.
    • The tier field is calculated using prompt[0]; prompt[1..4] may vary slightly in token length and should be re-tokenized if strict fixed bounds are required under a different text encoder (e.g., LFM2.5-230M or T5Gemma-2B).
  • Prompt Length Distribution: Captions are long descriptive paragraphs (typically 75–227 tokens). Models trained exclusively on this corpus expect descriptive prompts at inference time or LLM prompt rewriting for short inputs (i1 Finding 7).

Licensing & Attribution

The hf-dataset-toolkit curation scripts are licensed under Apache-2.0. The underlying images and captions remain subject to the respective licenses and terms of use of their upstream providers (FLUX-Reason-6M, GPT-Edit-Simpler, Midjourney-v6-Recap, Pexels, TextAtlas5M, and zlab-princeton/i1-captions). Users are responsible for complying with all upstream licensing terms when redistributing or training commercial models.


References

  • Asuka-FM: Asuka-FM: A System-Architectural Co-Design Framework for Efficient Text-to-Image Pretraining on Consumer Hardware project page.
  • i1: Zeng et al., i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models, arXiv:2606.11289 (2026).
  • i1 Synthetic Captions: zlab-princeton/i1-captions.
  • TextAtlas5M: Wang et al., TextAtlas5M: A Large-scale Dataset for Dense Text Image Generation, arXiv:2502.07870 (2025).
Downloads last month
2,737

Models trained or fine-tuned on aipracticecafe/i1-dataset

Papers for aipracticecafe/i1-dataset