Tag

MONET

Provides 104.9M curated image–text pairs with precomputed embeddings, structured annotations and pre-encoded VAE latents for text-to-image pretraining and retrieval. Combines filtered web sources and synthetic samples with multi-model re-captioning, deduplication and safety filters; Apache-2.0.

PiD — Pixel Diffusion Decoder

Converts latent representations into high-resolution images by using a conditional pixel-space diffusion decoder that merges decoding and upsampling into a single generative step. Released checkpoints are 4-step distilled (2k and 2kto4k variants) and pair with specific VAE/encoder weights; license restricts use to non-commercial research.

SenseNova-SI-8M

Curated multimodal training corpus for spatial intelligence: ~8.16M QA-style samples paired with ~2.72M unique images (≈1.1 TB). Provides JSONL annotations, a 1,000-sample preview, and 52 independent image archives — used to train SenseNova-SI models.

NuExtract3

Unified 4B vision-language model for document understanding that converts images or text into template-driven structured JSON or clean Markdown. Key features: multimodal inputs (image+text), template-based extraction, reasoning vs non-reasoning modes, and vLLM/OpenAI-compatible deployment for OCR, invoice/forms extraction, and RAG preprocessing.

VisCoR_Contrast (VisCoR-55K Contrastive Pairs)

Provides ~85K contrastive visual question–answer pairs where each example contains an anchor and a matched counterpart (image, question, answer). Pairs span General, Reasoning, Math, Graph/Chart and OCR categories to help train and evaluate fine‑grained, faithful visual reasoning in VLMs.

Pixal3D

Generates high-fidelity 3D assets from a single image by back-projecting pixel-aligned features into 3D, preserving fine geometry and PBR textures; includes inference code and a Hugging Face demo—best suited for single-view object reconstruction.

Thinking with Visual Primitives

Provides the dataset and accompanying technical report for a DeepSeek project that interleaves spatial markers (points and boxes) into multimodal LLM reasoning. Includes a public subset of data and benchmarks under an MIT license; model weights are not included.

L2P: Unlocking Latent Potential for Pixel Generation

Transforms pretrained latent-diffusion priors into pixel-space diffusion models by removing the VAE and training shallow pixel layers on LDM-generated synthetic images — enabling fast convergence, native 4K output, and low-data training on 8 GPUs.

stanford-vision-lab/gpic

Provides paired images and English captions for vision–language research, curated by Stanford Vision Lab and hosted on Hugging Face; useful for training and evaluating multimodal models and reproducing related research.

laion/BVD-V-55M

Provides 55 million scene-level video clips (each with captions, language labels, and timestamps) extracted from an 80M-video, 10-million-hour raw pool to support multimodal pre-training across video, audio, and frames. Access is gated for academic/non-commercial research.

Alibaba-YuFeng/MMA-SafetyBench

A small image-folder dataset for multimodal/vision model safety benchmarking, containing under 1,000 curated images with annotations to exercise safety-related model behaviours; licensed CC BY 4.0 and hosted on HuggingFace.

Stera-10M

Open egocentric multimodal dataset for embodied AI and robot learning captured on commodity iPhone Pro: ~200 hours and ~10M RGB frames with LiDAR depth, ARKit 6‑DoF poses, IMU, two‑hand MANO mocap, room meshes, and hierarchical action captions.