CoolFace
Datasetpublic

dingshizhe/vgtae-minit2i-sft-latents

MiniT2I SFT-set latents encoded with VGT-AE (448px) The text-to-image SFT mixture, pre-encoded into the VGT-AE latent space. Companion to dingshizhe/vgtae-cc12m-latents, which covers the pretraining corpus; same encoder, same settings, same layout. These are not DC-AE latents. VGT-AE is a hybrid codec — a fine-tuned Qwen2.5-VL ViT as the encoder, a DC-AE decoder at sampling time. The tensor shape coincides with DC-AE f32c32 (32, 16, 16), but the space is different.… See the full description on the dataset page: https://huggingface.co/datasets/dingshizhe/vgtae-minit2i-sft-latents.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face
0likes19downloads
Dataset Card

MiniT2I SFT-set latents encoded with VGT-AE (448px)

The text-to-image SFT mixture, pre-encoded into the VGT-AE latent space. Companion to `dingshizhe/vgtae-cc12m-latents`, which covers the pretraining corpus; same encoder, same settings, same layout.

These are not DC-AE latents. VGT-AE is a hybrid codec — a fine-tuned Qwen2.5-VL ViT as the encoder, a DC-AE decoder at sampling time. The tensor shape coincides with DC-AE f32c32 (32, 16, 16), but the space is different.

Contents

Two subsets, each a set of WebDataset tars aligned shard-for-shard with the source:

subsetshardssamples
blip3_ft60k/1158,859
dalle3/1419,024
total2577,883
blip3_ft60k/journeyDB.tar
  <key>.npy   float16, shape (32, 16, 16)
  <key>.txt   caption

blip3_ft60k is itself a mixture — its per-source shards are dalle3, geneval_train, human_gestures, journeyDB, mscoco_human, object_1/2, occupation_1/2, text_1/2. Note that the shard named dalle3.tar inside blip3_ft60k/ is not the same data as the top-level dalle3/ subset.

Encoder settings

EncoderVGTAE_Qwen25VL (Qwen2.5-VL-3B ViT → 3× ResBlock → down_mlp)
Input resolution448 px (bicubic resize on short edge, center crop, normalized to [-1, 1])
Stride28 → 16×16 latent grid = 256 tokens
embed_dim32
encoder_normTrue — channel-wise layer_norm_2d, so every latent has std == 1.0, mean == 0
Scale / shiftapplied — latents are stored scaled, do not scale again
Stochastic perturbationoff (encoder in eval() mode)
Storage dtypefloat16

Integrity

Every output tar was walked end to end after encoding: 0 unreadable or truncated shards, and the per-shard .npy counts sum to 77,883 — exactly the number of images in the source tars, so nothing was dropped.

Usage

python
import io, tarfile, numpy as np

with tarfile.open("blip3_ft60k/journeyDB.tar") as tar:
    for m in tar:
        if m.name.endswith(".npy"):
            z = np.load(io.BytesIO(tar.extractfile(m).read()))  # (32, 16, 16) fp16