Recraft V3

Om Recraft AI

Grundlagt 2022 · London, UK

Recraft AI is a generative-AI startup founded in 2022 in London by Anna Veronika Dorogush (CEO, former Yandex ML lead and CatBoost project lead) and other ex-Yandex engineers. The company raised a $12M Series A led by Khosla Ventures in 2024 and a follow-on round in early 2025, focusing on AI tools for designers, brand teams and marketing agencies. Recraft's main product is a design-focused web app and its proprietary Recraft V3 model (released October 2024), which topped the Artificial Analysis Image Arena leaderboard at launch, beating Midjourney v6.1, FLUX 1.1 [pro], DALL-E 3 and Ideogram 2.0 on human preference and prompt adherence. Recraft is positioned as a commercial design platform with strong typography, brand-style consistency and vector-output capabilities.

Arkitektur

Diffusion model with long-prompt language conditioning (proprietary)

Recraft V3 (codenamed 'red panda' during public benchmarking) is a closed-source text-to-image diffusion model from Recraft AI, released October 2024. Recraft has not published a technical report, but the model is widely reported to use a Diffusion Transformer backbone with a long-context LLM-based text encoder (handling up to ~1000-token prompts) and was trained with a heavy focus on graphic design, typography, brand consistency and aspect-ratio versatility. The model topped the Artificial Analysis text-to-image arena at launch (Elo ~1172) ahead of Midjourney v6.1 and FLUX 1.1 [pro]. Recraft also ships specialised variants — Recraft V3 Realistic (photography) and Recraft V3 SVG (vector graphics output) — built on the same underlying backbone. Outputs include support for long prompt understanding, in-image text up to several lines, brand kits (logos, palettes, fonts) and high-resolution rendering up to 2048x2048.

Parametre

Undisclosed

Kontekst

1.024 tokens

Funktioner

• #1 on Artificial Analysis text-to-image arena at launch
• Excellent typography and long in-image text
• Long-prompt understanding (~1000 tokens)
• Brand kits: logos, palettes, fonts applied across generations
• Multiple specialised variants (Realistic, SVG)
• Resolutions up to 2048x2048
• Available via Recraft web app and API
• Best for: brand-consistent ad creatives, posters, designed assets with text, agency workflows.

Træning og licens

Trained on a large licensed and publicly available image-text dataset with extensive design-focused augmentation. Exact composition is not disclosed.

Licens: Proprietary commercial license via the Recraft web app and API. Commercial use permitted under Recraft Terms of Service.

Sikkerhedstests: Internal safety filters; refuses CSAM, NCII, named persons and IP-protected characters. No public red-team report.