EchoVision-45K

Training data for Echo-Vision-1 — a two-stage coarse-to-fine rectified-flow text-to-image model trained from scratch on a single consumer GPU.

~45,000 image–caption pairs · max 768 px · machine-generated captions · copyright-filtered · compressed JPEG

---

Dataset Summary

EchoVision-45K is the curated dataset used to train Echo-Vision-1, a cascade latent-diffusion model (draft 384 px → final 768 px). It consists of square-cropped web images paired with detailed machine-generated captions.

• Images: max side 768 px, optimized JPEG (quality 85). Larger originals were downscaled because the model never consumes more than 768 px.
• Captions: generated by Florence-2-large (<MORE_DETAILED_CAPTION>), with ~7 k high-quality single-sentence captions from Qwen2-VL-2B-Instruct and BLIP fallbacks.
• Safety: a two-layer copyright filter (deterministic franchise/brand/watermark text blacklist + CLIP zero-shot visual flagging) removed logos, watermarks, and licensed-franchise content.

Usage

from datasets import load_dataset

ds = load_dataset("Maxilicious20/echo-vision-45k")
sample = ds["train"][0]
sample["image"]     # PIL.Image  (<=768px)
sample["caption"]   # str

Training-style transform:

from torchvision import transforms
tf = transforms.Compose([
    transforms.Resize((768, 768)),
    transforms.ToTensor(),
    transforms.Normalize([0.5], [0.5]),
])
x = tf(sample["image"])   # -> [3,768,768] in [-1,1]

Dataset Structure

• Config: imagefolder (auto-detected)
• Features: image (Image), caption (string)
• Split: train (~45 k examples; exact count in the Files/Viewer)

Data Collection & Processing

1. Collection: web images downloaded with associated queries.
2. Captioning: Florence-2-large / Qwen2-VL-2B / BLIP.
3. Filtering: resolution & sharpness (Laplacian) pre-filter.
4. Copyright removal: text blacklist (franchises, brands, "watermark", "fan art", ©/™/® …) + CLIP zero-shot visual flagging → flagged pairs deleted.
5. Compression: downscale to ≤768 px + JPEG q85 (no training-relevant information lost).

Limitations & Biases

• Web-scraped imagery inherits the biases of the open web (demographic, cultural, aesthetic).
• Machine captions can be noisy or incomplete.
• Despite filtering, residual copyrighted or identifiable content may remain; this dataset is intended for research only.

License & Third-Party Content

The dataset's original contributions — captions, metadata, curation/filtering code, and the compiled arrangement — are released under the Apache License 2.0.

The underlying images were collected from public web sources and remain the property of their respective copyright holders. The Apache-2.0 license does not grant rights to these third-party images. This dataset is provided for research only.

If you hold rights to any image and wish it removed, please open a discussion — it will be taken down promptly.

Citation

@misc{echovision45k,
  title  = {EchoVision-45K: A Copyright-Filtered Image--Caption Dataset
            for Text-to-Image Diffusion Research},
  author = {<Your Name>},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Maxilicious20/echo-vision-45k}}
}

---

Companion model: Echo-Vision-1 (Apache-2.0) · Built on a single RTX 4060.