minit2i_dcae — CC12M pre-encoded to DC-AE f32c32 latents (512 px)

CC12M images encoded once, offline, with the stock DC-AE mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers, so that latent-space training does not pay for the VAE forward pass every step.

Shards	1097 WebDataset .tar
Samples	6,836,022
Size	118.4 GiB (127.2 GB)
Latent	[32, 16, 16] float16 (512 px ÷ 32)
Sequence length	256 tokens at patch size 1
scaling_factor	0.41407 (already applied)
E[z²]	≈ 1.07

Layout

Each shard mirrors one source CC12M shard 1:1 and carries two members per sample:

00042.tar
  000412345.npy    float16 [32,16,16], numpy .npy, scaling_factor applied
  000412345.txt    caption (LLaVA-recaptioned), verbatim

Keys are the 9-digit global row index from the CC12M metadata, so a key identifies the same image across shards and metadata.

Loading

import io, numpy as np, webdataset as wds

def decode(sample):
    return np.load(io.BytesIO(sample["npy"])), sample["txt"].decode()

ds = (wds.WebDataset("https://huggingface.co/datasets/dingshizhe/minit2i_dcae/resolve/main/{00000..01096}.tar")
        .map(decode))
z, caption = next(iter(ds))          # z: float16 [32,16,16]

Back to pixels:

import torch
from diffusers.models import AutoencoderDC

ae = AutoencoderDC.from_pretrained("mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers").cuda().eval()
zt = torch.from_numpy(z).float().cuda()[None] / ae.config.scaling_factor
with torch.no_grad():
    img = ae.decode(zt).sample          # [1,3,512,512] in [-1,1]

How it was produced

Source JPEGs were already 512×512 (img2dataset applied center_crop at download time), so the pipeline is:

center square crop (no-op on square input) -> resize 512 -> /255 -> *2-1
  -> HWC→CHW -> DC-AE encode -> * scaling_factor -> float16

Because the sources are square, there is no random-crop augmentation to freeze — the precomputed latents are exactly what an on-the-fly encoder would produce from the same shards.

Precision

Encoding ran in fp16 and latents are stored in fp16. Measured on 64 held-out images against an fp32 reference encode:

Latent relative L2 error	0.482 %
PSNR, decode(shipped) vs decode(fp32)	61.45 dB
PSNR, decode(fp32) vs original pixels	24.44 dB
PSNR, decode(shipped) vs original pixels	24.44 dB

The error introduced by the precision choices sits ~37 dB below DC-AE's own reconstruction floor, i.e. roughly 5000× smaller in power, and is invisible in pixel space. fp16 was chosen over bf16 deliberately: at identical throughput it deviates 6× less (0.26 % vs 1.65 % mean, measured at encode time), because |z| ≤ 8.8 sits far inside fp16 range while fp16 carries 10 mantissa bits against bf16's 7.

Note on shard 00987

The source shard 00987 was left truncated by an interrupted img2dataset run (.tar cut mid-member, .parquet footer missing, _stats.json empty). It was repaired before encoding: the 6233 intact samples were kept, the half-written one was re-fetched, and every candidate URL the shard did not already hold was re-attempted, recovering 733 more images. That shard therefore holds 6967 samples, slightly above its neighbours (6220–6480) — the extra ones are CC12M images whose original download had been rejected with HTTP 403 by CDNs that refuse img2dataset's default user agent. Re-fetched samples are marked "refetched": true in the source shard's .json members.

Caveats

• 6.84 M samples, not CC12M's full 10.97 M: about 38 % of the URLs were dead (404/403/DNS) at download time in July 2026.
• Captions are the LLaVA recaptions shipped with the source metadata, not the original alt-text.
• Latents are frozen to 512 px / [32,16,16]. Training at another resolution requires re-encoding from pixels.