dc-ai/DC-Gen-Data

DC-Gen Data

This repository contains example training data and processed evaluation data for DC-Gen. The training examples are small subsets for exercising data preparation and model adaptation workflows; they are not the complete datasets used to train the released models. The evaluation inputs are prepared test data for Wan I2V benchmarking.

Contents

Subfolder	Contents	Intended use
train_example/flux/	6 original TAR shards, including an align/ subset	FLUX RGB metadata preparation, teacher/student latent extraction, and adaptation training examples
train_example/zimage/	5 original TAR shards, including an align/ subset	Z-Image RGB metadata preparation, teacher/student latent extraction, and adaptation training examples
train_example/qwen-image-edit/qwen_image_gen/	1 TAR of Qwen-Image-Edit generated examples	Embedding alignment and patch/head training examples
train_example/qwen-image-edit/single_turn_processed/	1 TAR of processed Pico-Banana examples	Subsequent image-editing training examples
train_example/wan/fusionX_480p/	1 original TAR under align/shift_2/	Wan/DC-AE-V latent extraction and T2V/I2V adaptation training examples
eval/wan-i2v/	355 images at 832x480 plus VBench_i2v_extended_full_info.json	Prepared VBench I2V test inputs with extended prompts

Original TAR filenames, contents, and internal directory structure are preserved. Metadata indexes and latent archives are generated locally using the project recipes. Model weights are distributed separately.

Download only the required subfolder

Install huggingface_hub if needed. Use HfFileSystem.get with a trailing / on the remote subfolder to download its contents directly into the directory already used by your training or evaluation commands. Only the selected subfolder is downloaded; the remote train_example/ or eval/ prefix is not added to the local destination.

For the processed Wan I2V evaluation data:

python - <<'PY_DOWNLOAD'
from huggingface_hub import HfFileSystem

HfFileSystem().get(
    "datasets/dc-ai/DC-Gen-Data/eval/wan-i2v/",
    "assets/data/vbench",
    recursive=True,
)
PY_DOWNLOAD

This creates assets/data/vbench/VBench_i2v_extended_full_info.json and assets/data/vbench/vbench_i2v_imgs/832-480/, matching the existing VBenchImagePrompt defaults. No trainer path overrides are needed.

For the Wan training examples:

python - <<'PY_DOWNLOAD'
from huggingface_hub import HfFileSystem

HfFileSystem().get(
    "datasets/dc-ai/DC-Gen-Data/train_example/wan/fusionX_480p/",
    "assets/dataset/fusionX_480p",
    recursive=True,
)
PY_DOWNLOAD

For other training examples, choose the same local directory already used by the project's data preparation commands:

Remote subfolder (append to datasets/dc-ai/DC-Gen-Data/)	Local destination
train_example/flux/	Your FLUX RGB TAR directory (<path/to/rgb/tar/shards>)
train_example/zimage/	Your Z-Image RGB TAR directory (<path/to/rgb/tar/shards>)
train_example/qwen-image-edit/single_turn_processed/	$PICO_BANANA_DATA_PATH/single_turn_processed
train_example/qwen-image-edit/qwen_image_gen/	$PICO_BANANA_DATA_PATH/qwen_image_gen
train_example/wan/fusionX_480p/	assets/dataset/fusionX_480p
eval/wan-i2v/	assets/data/vbench

Keep the trailing / on the remote subfolder so the selected contents map directly into the destination, including nested directories such as align/. Download the Qwen subfolders separately when only one is needed.

Training follows the sequence RGB TARs → metadata indexes → latent archives → latent metadata → training. Ordinary manual-prompt inference uses model weights and user inputs and does not require these training examples or benchmark inputs.

Sources and evaluation processing

The training examples are migrated from the DC-Gen project's existing shared data folders. single_turn_processed contains processed Pico-Banana examples; qwen_image_gen contains the corresponding project's Qwen-generated training examples. This release preserves the supplied sample archives rather than generating additional training data.

The I2V test images are derived from VBench's I2V benchmark, prepared at a 26:15 aspect ratio and 832x480 resolution. The accompanying JSON contains the project's extended prompts. This repository distributes those processed inputs; it does not include the official origin.zip, crop.zip, or other VBench release archives. Refer to the upstream projects for original data terms and attribution.

The JSON is the version paired with the project's shared prepared images. The separate VBench-I2V-Prompt-Extended repository is retained unchanged.

checksums.sha256 records SHA-256 checksums for every data file in this release.