Recraft AI
Last reviewed
Sources
20 citations
Review status
Source-backed
Revision
v6 ยท 4,705 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
20 citations
Review status
Source-backed
Revision
v6 ยท 4,705 words
Add missing citations, update stale details, or suggest a clearer explanation.
Recraft AI is an AI image generation platform built for professional designers, creative teams, and brand-focused workflows. Founded in 2022 by machine learning scientist Anna Veronika Dorogush, the company produces both raster and vector graphics from text prompts, a capability that distinguishes it from most competitors in the generative AI space.[2] Recraft's proprietary V3 model (codenamed "Red Panda") reached the top position on the Artificial Analysis Text-to-Image Leaderboard in October 2024 with an ELO rating of 1172 and a 72% win rate, surpassing Midjourney, DALL-E 3, Stable Diffusion, and Ideogram.[5] As of mid-2025, Recraft reported over 7 million registered users across more than 200 countries, and the company has raised $42 million in total funding.
The company headquarters are in San Francisco, California. Recraft positions itself as a design-first AI platform rather than a general-purpose image generator, focusing on the daily needs of professional designers who require brand consistency, precise layout control, and production-ready output in formats such as SVG, PNG, JPG, PDF, TIFF (up to 300 DPI CMYK), and Lottie.[11] Clients include Netflix, HubSpot, Ogilvy, Amazon, NVIDIA, Salesforce, and Uber.
Anna Veronika Dorogush founded Recraft in 2022 after building a career in machine learning research across some of the largest technology firms in the world.[2]
She graduated from the Faculty of Computational Mathematics and Cybernetics at Lomonosov Moscow State University with a Master of Science in Applied Mathematics and Computer Science, completing the program from 2005 to 2010.[2] She also attended the Yandex School of Data Analysis. Before entering technology full-time, she worked as a professional model while completing her mathematics degree. On that period she has said: "The biggest lesson from that time was that grinding isn't everything."[2]
Her professional career began at ABBYY in 2011 as a software engineer in the Morphology team, where she worked on word processing features including spellcheckers and automatic spelling correction.[2] She moved to Microsoft in 2012 and to Google in 2014, where she worked as a software engineer in the natural language processing group.[2]
In 2016, Dorogush joined Yandex (often described as Russia's Google) as an individual contributor and Head of ML Tools, where she built the team responsible for developing machine learning libraries and tools.[2] Her most consequential technical achievement at Yandex was co-creating CatBoost, an open-source gradient boosting library released in 2017.[19]
CatBoost addressed a fundamental limitation of previous gradient boosting frameworks: the need for extensive manual preprocessing of categorical features. The library introduced a permutation-driven approach to encoding categorical variables that significantly reduced the risk of target leakage and improved out-of-the-box performance on real-world datasets.[15] Dorogush, along with Vasily Ershov and Andrey Gulin, authored the foundational paper "CatBoost: gradient boosting with categorical features support," which was presented at the Workshop on ML Systems at NeurIPS 2017.[15] The academic publications related to CatBoost accumulated over 2,266 citations.
InfoWorld magazine recognized CatBoost as one of the best machine learning tools of 2017, alongside TensorFlow and PyTorch.[19] The library is used in production by organizations including CERN, Cloudflare, Careem, and JetBrains, and powers ranking, recommendation, and forecasting systems within Yandex itself.[19]
Dorogush left Yandex to found Recraft in 2022.[2] The jump from gradient boosting libraries to generative image models reflects her broader interest in building foundational machine learning tools rather than consumer applications, applied this time to visual design.
Recraft was founded in 2022 with the goal of building an AI design tool purpose-built for professional creative workflows. The company operated in stealth during its initial development phase. On May 31, 2023, Recraft emerged from stealth with a public launch on Product Hunt, showcasing its ability to generate vector graphics from text prompts. This was a novel capability at the time; existing AI image generators like Midjourney and Stable Diffusion produced only raster images.
In January 2024, Recraft raised a $12 million Series A round led by Khosla Ventures.[4] This funding enabled the company to expand its model development and begin building out its enterprise offering. During 2024, Recraft released its V2 model and continued iterating on its vector generation technology. The V2 model brought improved image quality, better prompt adherence, and expanded style control options compared to the initial V1 release.
On October 28, 2024, an anonymous model codenamed "Red Panda" appeared at the top of the Artificial Analysis Text-to-Image Model Leaderboard, a crowdsourced benchmark hosted on Hugging Face.[5] The model surpassed all existing image generation systems, including Midjourney v6.1, Ideogram v2, and OpenAI's DALL-E 3 HD.[5] The identity of Red Panda's creator was initially unknown, generating significant speculation in the AI community.[6]
Recraft subsequently confirmed that Red Panda was its Recraft V3 model.[6] The reveal drew attention from major technology publications including TechCrunch and Tom's Guide.[1] TechCrunch covered the story under the headline "A stealth AI model beat DALL-E and Midjourney on a popular benchmark."[1] Tom's Guide described it as a potential threat to Midjourney's dominance.[6] Recraft V3 maintained its top position on the leaderboard for over five consecutive months.[6]
On May 5, 2025, Recraft announced a $30 million Series B funding round led by Accel, with participation from Madrona Ventures, Khosla Ventures, Nat Friedman (former GitHub CEO), Elad Gil, and RTP Global.[4] At the time of the announcement, the company reported over 4 million registered users and more than $5 million in ARR.[4] The company described itself as being in "hyper-growth mode," citing 10x expansion in two years and a 700% increase in user adoption over the prior year.[4] Revenue grew to $8.4 million by September 2025.[12]
In July 2025, Recraft joined the Model Context Protocol (MCP) ecosystem, allowing its image generation capabilities to be called directly from AI tools such as Claude, Cursor, and other AI agents.[8] This shifted Recraft from being a standalone design platform to a flexible creative engine accessible from within other AI-powered development environments. The MCP server is open-source and available on GitHub at recraft-ai/mcp-recraft-server.[8]
In December 2025, Recraft introduced Agentic Mode, an LLM-powered conversational interface that allows designers to generate and refine visuals through natural language conversation rather than manual prompt rewriting.[9] The system understands context across multiple turns, tracks conversation history, and can suggest adjustments, apply styles, or select models automatically.[9]
On February 17, 2026, Recraft released V4, described as a "ground-up rebuild" focused on design taste.[7] The V4 model was developed in close collaboration with working designers and tuned to make intentional aesthetic decisions about composition, color relationships, lighting mood, and material detail.[7] V4 is available in four configurations: Standard (1024x1024), Pro (2048x2048), Vector (SVG output), and Vector Pro (high-resolution SVG).[7] All versions are accessible to every user, including those on the free plan.[7]
Recraft has raised a total of $42 million across two funding rounds.
| Round | Date | Amount | Lead Investor | Other Investors |
|---|---|---|---|---|
| Series A | January 2024 | $12M | Khosla Ventures | Not disclosed |
| Series B | May 2025 | $30M | Accel | Madrona Ventures, Khosla Ventures, Nat Friedman, Elad Gil, RTP Global |
Madrona Ventures published a detailed investment thesis explaining their Series B participation, noting that Recraft was "purpose-built from day one" for professional designers rather than retrofitted from a general-purpose model.[3] The Madrona piece highlighted the company's unusual combination of model quality (top leaderboard ranking) and design-specific tooling as the core of their investment reasoning.[3]
Khosla Ventures led the Series A and continued into the Series B, a pattern that reflects conviction in the team's technical direction.[4] Nat Friedman's involvement as an angel investor is notable given his background as the former CEO of GitHub and longtime observer of developer and creative tool markets.[4]
Recraft has released four major model versions since its founding.
| Version | Release Date | Key Features |
|---|---|---|
| V1 | 2023 | Initial model released with the Product Hunt launch. Introduced vector graphics generation from text prompts. |
| V2 | 2024 | Improved image quality, better prompt adherence, and expanded style control options. |
| V3 (Red Panda) | October 2024 | 20-billion-parameter model. Reached #1 on the Artificial Analysis leaderboard with ELO 1172. First AI model to generate long-form text in images. Native SVG vector output. Median generation speed of approximately 7 seconds. |
| V4 | February 2026 | Ground-up rebuild focused on "design taste." Available in Standard (1024x1024), Pro (2048x2048), Vector, and Vector Pro variants. Improved composition, color, lighting, and typography integration. Developed in collaboration with professional designers. |
Recraft V3, codenamed "Red Panda," is a 20-billion-parameter text-to-image model released in October 2024.[5] Unlike many competing systems that build on variants of existing open-source architectures, Recraft built V3 from scratch. The model achieved a 72% win rate and ELO rating of 1172 on the Artificial Analysis leaderboard, outperforming Midjourney v6.1 (ELO 1093), FLUX1.1 Pro (ELO 1143), and DALL-E 3 HD (ELO 984) at time of release.[5]
The V3 model was notable for being the first AI image generator to reliably render long-form text within images, including full sentences and paragraphs rather than just individual words. It also demonstrated improved anatomical accuracy, including correct finger counts and balanced body proportions, addressing failures that had been persistent in earlier diffusion-based systems.
Generation speed for V3 runs at approximately 7 seconds per image, which the company reported as roughly twice as fast as DALL-E 3 at the time of V3's release.
Recraft V4, released February 17, 2026, is a ground-up rebuild with a specific design philosophy: make intentional aesthetic choices about composition, color, light, and material the way a skilled art director would, rather than averaging toward the median of training data.[7]
V4 treats text as a structural component of composition rather than a purely decorative overlay, enabling it to resolve spatial relationships between typography and other visual elements with greater precision.[13] The model supports prompts up to 10,000 characters.[17]
V4 comes in four variants:
| Variant | Resolution | Processing Time | API Cost | Use Case |
|---|---|---|---|---|
| V4 Standard | ~1024px | ~10 seconds | $0.04/image | Iteration, everyday work |
| V4 Pro | ~2048px | ~28 seconds | $0.25/image | Print-ready assets, large-scale use |
| V4 Vector | Editable SVG | ~15 seconds | $0.08/image | Brand assets, icons, illustrations |
| V4 Vector Pro | High-res SVG | ~30 seconds | $0.30/image | High-detail vector production |
All four variants are available to every user, including those on the free plan.[7]
Recraft's ability to generate true native SVG vector graphics is its most technically distinctive feature and the one that separates it most clearly from competing platforms.
Most AI image generators, including Midjourney, DALL-E, and Stable Diffusion, produce raster images: grids of pixels that become blurry or pixelated when enlarged beyond their native resolution. Some tools offer post-generation vectorization, which traces raster output into approximate paths. The result is usually messy path structures with hundreds of overlapping anchor points, suitable only for background shapes at best.
Recraft's approach is different. The model generates SVG files natively, meaning the output contains actual paths, structured geometry, and clean composition from the start. These SVGs can be opened directly in Figma, Adobe Illustrator, Sketch, or any other vector editor. Paths are editable at the node level. Shapes are logically structured. Colors are applied as fills and strokes, not baked into pixels.
This makes Recraft the practical choice for workflows where the AI output will be further refined by a human designer: a logo concept that needs to be adjusted by a brand team, icon variants that need to be resized for different contexts, or an illustration that needs color adjustments to match a client's palette.
For V4, the Vector and Vector Pro variants extend this capability with improved path quality and finer structural coherence compared to V3's SVG output.[13]
Brand consistency is one of Recraft's core product focuses, and the platform has built a dedicated suite of style tools around this problem.[10]
The platform maintains a searchable library of over 100 curated visual styles, ranging from photorealistic photography to flat vector illustration, editorial photography, product renders, and typographic posters.[10] Users can search by keyword or filter by type to browse the available aesthetics.
Users can create custom styles by uploading reference images.[10] The platform supports two style models:
These custom styles can then be applied to any subsequent generation, producing outputs that match the visual identity of the reference without requiring the user to describe it in a text prompt. This is particularly useful for brand teams that have existing visual libraries but no standardized prompt language to describe them.
Recraft supports blending up to five styles simultaneously, with adjustable weighting between each source.[10] A designer can combine a clean sans-serif editorial style with a particular color palette from a second reference and a paper texture from a third, and the system will produce generations that synthesize all three influences. Additional refinement is possible through style-level prompts such as "centered layout" or "transparent plastic."
Saved styles can be shared across a team's workspace, which allows an organization to lock in a visual identity and propagate it across all members working on related projects.[10] This addresses the practical problem of creative teams where different designers, working independently, produce outputs with inconsistent aesthetics.
Beyond generation, Recraft's studio includes a range of image editing tools that allow designers to refine AI outputs or modify existing images.
| Tool | Function |
|---|---|
| Inpainting | Edit specific areas by painting a mask and providing new instructions |
| Outpainting | Extend an image beyond its original borders to add context or complete a cropped scene |
| Background removal | Automatically isolate subjects by removing backgrounds |
| Background replacement | Swap backgrounds with new generated or solid color backgrounds |
| Creative Upscale | Increase resolution while adding new detail (20 credits in the studio) |
| Crisp Upscale | Increase resolution while preserving existing detail closely |
| AI Eraser | Remove specific unwanted elements from an image |
| Modify Area | Make targeted changes to specific regions |
| Vectorization | Convert raster images to SVG vector format |
| AI Mockuper | Generate product mockups with AI assistance |
| AI Fine-Tuning | Customize model behavior for specific use cases |
The studio provides an infinite canvas where multiple images can be arranged, edited, and compared simultaneously. Real-time collaboration is supported for team accounts.
The Recraft API allows developers to integrate image generation, vector graphics creation, and image editing into their own applications.[11] It is a REST API with no GPU management or infrastructure setup required from the developer.
Supported operations include:
| Operation | Description |
|---|---|
| Image generation | Generate raster images from text prompts |
| Vector generation | Generate SVG vector graphics from text prompts |
| Image-to-image | Transform or restyle existing images |
| Prompt-based editing | Modify images using text instructions |
| Background removal | Remove backgrounds programmatically |
| Background replacement | Replace backgrounds via API |
| Inpainting and outpainting | Edit or extend images programmatically |
| Vectorization | Convert raster images to SVG |
| Style creation | Create custom styles via API for consistent generation |
| Batch jobs | Process multiple generation requests in bulk |
| Asynchronous processing | Queue long-running jobs and retrieve results later |
The API supports output in SVG, PNG, JPG, PDF, TIFF (up to 300 DPI CMYK), and Lottie formats.[11] The API also supports style control with exact RGB color specifications, custom background colors, and artistic style parameters.
In March 2025, Recraft significantly reduced API pricing across several operations. Creative Upscale dropped from $0.80 to $0.25 per image, Crisp Upscale from $0.04 to $0.004, and background removal from $0.04 to $0.01.[11]
Recraft's open-source MCP server (recraft-ai/mcp-recraft-server on GitHub) allows AI agents including Claude and Cursor to call Recraft's image generation and editing capabilities directly within their workflows.[8] Through MCP, agents can create and edit raster and vector images, apply custom styles, manipulate backgrounds, upscale images, and perform vectorization.[8]
Recraft operates on a freemium model with three user-facing tiers.[20]
| Plan | Monthly Price | Credits | Key Features |
|---|---|---|---|
| Free | $0 | 30 daily (renewable) | Basic editing tools, community styles library. No commercial license. All images are public and posted to the community gallery. |
| Pro | From $10/month (billed annually) | 1,000+ monthly | Private image generation, full commercial license, priority generation (up to 10 parallel jobs), Creative Upscale, access to external models. |
| Teams | From $55/month (billed annually) | 9,000+ monthly | Shared workspace, shared credit pool, everything in Pro, premium support, SSO options for enterprise security requirements. |
Different operations consume different amounts of credits within the studio:
| Action | Credits Used |
|---|---|
| Raster image generation | 1 credit |
| Vector image generation | 2 credits |
| Creative Upscale | 20 credits |
Monthly plan credits do not roll over between billing cycles.[20] Top-up credits purchased by paid users do not expire.[20]
The Recraft API uses a unit-based pricing model. Units are purchased in advance at $1.00 per 1,000 units and do not expire.[11] Per-operation costs for the V4 model family:
| Operation | Cost |
|---|---|
| V4 Standard raster generation | $0.04 (40 units) |
| V4 Pro raster generation | $0.25 (250 units) |
| V4 Vector SVG generation | $0.08 (80 units) |
| V4 Vector Pro SVG generation | $0.30 (300 units) |
| Vectorization | $0.01 (10 units) |
| Background removal | $0.01 (10 units) |
| Crisp Upscale | $0.004 (4 units) |
| Creative Upscale | $0.25 (250 units) |
| Region erasing | $0.002 (2 units) |
| Image variation | $0.04 (40 units) |
New API users receive free credits to test operations before committing to a package purchase.
Recraft serves over 7 million creators across more than 200 countries. Enterprise and notable clients include:
| Company | Industry | Use |
|---|---|---|
| Netflix | Entertainment | Design and marketing assets |
| HubSpot | Marketing software | Brand and campaign visuals |
| Ogilvy | Advertising agency | Client creative production |
| Amazon | E-commerce, technology | Product and marketing imagery |
| NVIDIA | Semiconductors, AI | Design and technical assets |
| Salesforce | Enterprise software | Marketing design |
| Uber | Ride-sharing, technology | Brand and design assets |
| Airbus | Aerospace | Technical and marketing visuals |
| Asana | Project management software | Marketing visuals |
The presence of major advertising agencies like Ogilvy alongside technology companies reflects Recraft's position as a tool with traction at both ends of the professional design spectrum: brand-building agencies and in-house design teams at large technology firms.
Recraft competes in the AI image generation space against several well-funded platforms, each with distinct strengths and target audiences.
| Platform | Developer | ELO Rating (Oct 2024) | Win Rate | Key Strengths | Key Weaknesses |
|---|---|---|---|---|---|
| Recraft V3 | Recraft | 1172 | 72% | Native SVG vector output, brand consistency, long-form text rendering, positioning control | Smaller ecosystem than established competitors |
| FLUX 1.1 Pro | Black Forest Labs | 1143 | 68% | Photorealistic output, open-source base models, strong for photography | No native vector output |
| Midjourney v6.1 | Midjourney, Inc. | 1093 | 64% | Artistic quality, cinematic aesthetics, large community | No vector export, text rendering inconsistent, public gallery raises privacy concerns |
| Ideogram v2 | Ideogram | 1102 | 63% | In-image text accuracy, poster and label design | Less suited for complex design workflows |
| Stable Diffusion 3 Large Turbo | Stability AI | 1084 | 61% | Open-source, full customization, local processing | Requires technical setup for optimal results |
| Adobe Firefly 3 | Adobe | 971 | N/A | Adobe Creative Cloud integration, licensed training data | Slowest generation speed (over 1 minute per prompt), lower ELO |
| DALL-E 3 HD | OpenAI | 984 | 51% | ChatGPT integration, conversational generation | No vector output, lower benchmark performance |
Midjourney and Recraft have emerged as the two most prominent AI image generation tools for professional creatives, but they serve meaningfully different purposes.[14] Midjourney excels at artistic vision and cinematic concepts. Its models are tuned for aesthetic expressiveness, and the platform's Discord-based community has shaped a distinct generative aesthetic that has influenced the broader field. However, Midjourney does not produce vector output, its text rendering is inconsistent, and its public gallery means generated images are visible to other users by default, which creates problems for clients whose work requires confidentiality.[14]
Recraft produces images with a cleaner, more design-appropriate aesthetic, particularly for brand assets, icons, and print materials.[14] Its generation speed is approximately 4.5 seconds compared to Midjourney's typical range, and its native SVG output makes it the only practical AI option for workflows that terminate in vector editing software.
A common professional workflow combines both platforms: Midjourney for conceptual ideation and mood exploration, Recraft for production-ready vector assets and brand-consistent deliverables.[14]
Adobe Firefly's competitive advantage is its integration into the Creative Cloud ecosystem.[14] Designers who already work in Photoshop and Illustrator can use Firefly's generative fill, background generation, and recoloring features without leaving their existing workflow. Firefly is also trained exclusively on licensed Adobe Stock content, which gives it a legally defensible position for commercial use that some enterprises require.[14]
Recraft's advantages over Firefly are generation quality (significantly higher ELO score), generation speed (Firefly often exceeds one minute per prompt), and native SVG output.[14] Recraft's API is also more accessible for developers building image generation into their own products. For teams not already embedded in the Adobe ecosystem, Recraft offers better standalone performance.
FLUX models from Black Forest Labs are particularly strong for photorealistic output and have a large open-source community around them. FLUX 1.1 Pro sits just below Recraft V3 on the Artificial Analysis leaderboard (ELO 1143 vs. 1172).[5] FLUX's primary use case is photorealistic photography-style content: product shots, portraits, and realistic scenes. Recraft's primary use case is design-oriented output: logos, icons, brand assets, and illustrations. The two tools target overlapping but distinct segments of the professional creative market.
Recraft is used across a range of professional design applications:
Recraft received substantial coverage following the October 2024 reveal that its V3 model was the anonymous "Red Panda" that had topped the Artificial Analysis leaderboard.[1] The surprise reveal generated attention because the model had outperformed all known competitors without any public announcement.[6]
TechCrunch covered the story under the headline "A stealth AI model beat DALL-E and Midjourney on a popular benchmark, its creator just landed $30M."[1] Tom's Guide ran a piece titled "Mystery of Red Panda is solved: AI company Recraft behind leading new image model," describing the result as a potential threat to Midjourney's market position.[6] Recraft V3 held the top leaderboard position for over five consecutive months after its release.[6]
Madrona Ventures' investment thesis blog post, published at the time of the Series B, drew attention from the design and venture communities.[3] The post described the company as having achieved a rare combination in AI: genuinely top benchmark performance on a rigorous third-party evaluation alongside purpose-built tooling for a specific professional audience.[3]
The company's growth metrics have been noted in funding coverage. Revenue reached $8.4 million by September 2025 with approximately 50 employees, a per-employee revenue figure that drew favorable comparison to other funded AI startups.[12]
Despite strong benchmark performance, Recraft has several practical limitations that users have noted: