---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: image
    dtype: image
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 1145721433301.688
    num_examples: 2810604
  download_size: 5473812704
  dataset_size: 1145721433301.688
language:
- en
task_categories:
- text-to-image
license: cc-by-nc-4.0
tags:
- diffusion
- controlnet
- image-generation
size_categories:
- 1M<n<10M
---

# Dataset Card for "MultiGen-20M_train"

This dataset is constructed from [UniControl](https://arxiv.org/abs/2305.11147) and is a large-scale dataset for conditional text-to-image generation.
It is notably used for training and evaluation in:

*   The paper [ControlNet++: Improving Conditional Controls with Efficient Consistency Feedback](https://huggingface.co/papers/2404.07987).
    Its GitHub repository: [https://github.com/liming-ai/ControlNet_Plus_Plus](https://github.com/liming-ai/ControlNet_Plus_Plus)
*   The paper **[Heeding the Inner Voice: Aligning ControlNet Training via Intermediate Features Feedback](https://huggingface.co/papers/2507.02321)** (InnerControl).
    *   Project page: [https://controlgenai.github.io/InnerControl/](https://controlgenai.github.io/InnerControl/)
    *   Code: [https://github.com/control/InnerControl](https://github.com/control/InnerControl)

The dataset consists of various control types like LineArt, Hed, and Depth, and is designed to improve spatial control over generated images in text-to-image diffusion models.

### Abstract of "Heeding the Inner Voice":

Despite significant progress in text-to-image diffusion models, achieving precise spatial control over generated outputs remains challenging. ControlNet addresses this by introducing an auxiliary conditioning module, while ControlNet++ further refines alignment through a cycle consistency loss applied only to the final denoising steps. However, this approach neglects intermediate generation stages, limiting its effectiveness. We propose InnerControl, a training strategy that enforces spatial consistency across all diffusion steps. Our method trains lightweight convolutional probes to reconstruct input control signals (e.g., edges, depth) from intermediate UNet features at every denoising step. These probes efficiently extract signals even from highly noisy latents, enabling pseudo ground truth controls for training. By minimizing the discrepancy between predicted and target conditions throughout the entire diffusion process, our alignment loss improves both control fidelity and generation quality. Combined with established techniques like ControlNet++, InnerControl achieves state-of-the-art performance across diverse conditioning methods (e.g., edges, depth).

![InnerControl Method Diagram](https://cdn-uploads.huggingface.co/production/uploads/6667e3d60a7f1d1cbb63cf4d/IEOsiqRC9pV_SgzmN9zos.png)

### Data structure

Folder `./data/data_part0` contains an example subset and folder `./data/full_data` contains the full data currently, 1.04M images in total. Under each `.zip` package, data is arranged as below:

```
- package_idx
--- package_idx.json # metadata
--- images
----- 00001.png
----- 00002.png
...
```

Each sample is a `2x2` image grid at the resolution of `1024x1024`, we count the data sample as `ONE` sample, allowing for more space and diversity for randomly choosing different sample within the 4 images at training time.

`metadata.json` contains the metadata for each sample, where we show an example as below:

```json
# Metadata structure
[
    {
      "idx": ..., # index
      "image_path": "", # path to the image
      "features": {
            "attributes": [{"attribute type": "..."}, {"attribute type": "..."}], # attribute types and the specific description in this sample.
            "subject": "..." # subject name
        },
      "prompt": "...", # the prompt used for image generation
    },
  ...
]

# An example
[
    {
      "idx": 0,
      "image_path": "0/00000.png",
      "features": {
            "attributes": [
                {
                    "lighting": "1@hard light, highlighting texture, sharp shadow"
                },
                {
                    "color": "30@soft lavender, mint green, pale peach, and baby blue"
                }
            ],
            "subject": "introspective sports portrait"
        },
       "prompt": "soft lavender, mint green, pale peach, and baby blue, hard light, highlighting texture, sharp shadow, introspective sports portrait",
    },
  ...
]
```

### Code and supporting files

**Attributes and Subjects**

`./code/attributes_and_subjects.json` contains the attribute and subject dictionaries.

**Range-sensitive filtering**

`./code/range_sensitive_filter.json` contains our meta data for the filter, and `./code/data_filter.py` converts it into a format that can be used in the dataloader.

**Data Loader**

`./code/dataloader.py` provides an example in loading the data into image pairs with the filter and balanced resampling adopted.