DALL-E 3

Sobre OpenAI

Fundado em 2015 · San Francisco, USA

OpenAI was founded in December 2015 by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, Wojciech Zaremba and John Schulman as a non-profit research lab with a $1B funding pledge. Restructured as a capped-profit company in 2019, OpenAI introduced its DALL-E text-to-image family beginning with DALL-E (Jan 2021), DALL-E 2 (April 2022) and DALL-E 3 (Sept-Oct 2023). DALL-E 3 was developed in close collaboration with Microsoft and built into ChatGPT and Bing Image Creator. The DALL-E team includes Aditya Ramesh, Prafulla Dhariwal, Mark Chen and Gabriel Goh. OpenAI is led by CEO Sam Altman and is majority-funded by Microsoft (>$13B invested), with a 2025 valuation north of $150B.

Arquitetura

Latent diffusion model with LLM-based prompt rewriter (cascade)

DALL-E 3 is a closed-source latent diffusion text-to-image system built on top of a large pretrained image autoencoder. Its core innovation, described in the technical report 'Improving Image Generation with Better Captions' (Betker et al. 2023), is dataset recaptioning: a fine-tuned vision-language captioner was used to relabel large portions of the training corpus with rich, detailed synthetic captions that describe content, style and composition far more thoroughly than the noisy alt-text used by earlier models. DALL-E 3 is also tightly integrated with GPT-4: at inference, a prompt-rewriter LLM expands the user's short prompt into a longer, more descriptive caption before it is fed to the diffusion backbone, which substantially improves prompt adherence. The diffusion backbone uses a U-Net with cross-attention conditioning and is trained with the standard noise prediction objective; sampling typically uses 50 DDIM/DPM-Solver steps. Safety filters and a multi-stage moderation pipeline (input and output) are applied, and the model refuses requests for named living people, copyrighted characters and explicit content.

Parâmetros

Undisclosed (estimated multi-billion parameters in U-Net plus rewriter LLM)

Contexto

4.000 tokens

Capacidades

• Excellent prompt adherence due to GPT-4-based prompt rewriting
• High-quality, detailed images at 1024x1024, 1792x1024 and 1024x1792 resolutions
• Strong typography and short-text rendering inside images
• Coherent multi-object scenes with correct spatial relations
• Style control via natural-language style descriptors
• Built-in safety: refuses named real people, IP characters and unsafe content
• Tight ChatGPT integration including iterative refinement chat
• C2PA content credentials embedded in outputs
• Best for: marketing creatives, blog illustrations, concept art, presentation graphics.

Treinamento & licença

Trained on a large, licensed and publicly available image-text corpus with extensive synthetic recaptioning by a fine-tuned vision-language model. Exact data sources are not disclosed. OpenAI emphasises filtering for explicit content, violent imagery and personally identifiable information.

Licença: Proprietary commercial license via OpenAI API, ChatGPT Plus/Team/Enterprise and Microsoft Bing/Copilot.

Testes de segurança: Extensive pre-release red-teaming covering bias, deepfakes, CSAM, copyright and political imagery, plus a Preparedness Framework evaluation. Outputs are watermarked with C2PA content credentials.