CtrlK
Recraft V3 is a text-to-image model developed by Recraft AI, released on October 30, 2024. Built on a 20-billion-parameter architecture trained from scratch, it debuted under the internal codename "red_panda" and quickly took the top position on the Artificial Analysis Text-to-Image Arena leaderboard, surpassing models from Black Forest Labs, Midjourney, and OpenAI. It is the first widely available image generation model to produce native scalable vector graphics (SVG) output and remains, as of 2026, the only model capable of rendering long passages of text at specified positions within an image. A successor model, Recraft V4, was released in February 2026.
Recraft AI was founded in 2022 by Anna Veronika Dorogush, a machine learning scientist who had previously co-created the CatBoost gradient-boosting library at Yandex and held roles at Google and Microsoft. The company is headquartered in San Francisco and emerged from stealth on May 31, 2023, when it publicly demonstrated vector graphics generation on Product Hunt.
Recraft's founding premise was that existing text-to-image models were not built for professional design work. Tools like DALL-E and Stable Diffusion produced raster images with frequent anatomical errors, no vector support, and limited ability to render readable text inside images. Dorogush set out to build a model trained from the ground up around design requirements rather than retrofitting an existing architecture.
The company raised a $12 million Series A from Khosla Ventures in January 2024. In May 2025, it closed a $30 million Series B led by Accel with participation from Madrona Ventures and existing investors including Nat Friedman and Elad Gil. At the time of the Series B announcement, Recraft reported more than 4 million registered users and over $5 million in annual recurring revenue, with 700% user growth over the prior year.
Recraft's first publicly available model was released in 2023 alongside the company's Product Hunt launch. In March 2024, the company released Recraft V2 (also called Recraft 20B), described internally as the first model built from scratch to professional design standards. V2 introduced the 20-billion-parameter architecture that V3 would later inherit and expand upon, and demonstrated that a custom-trained model could match or exceed the output of fine-tuned variants of Stable Diffusion and DALL-E 3 on benchmark tasks involving perspective accuracy, quantity representation, and object positioning.
Recraft V3 was released publicly on October 30, 2024. In the days before the official announcement, a mysterious entry labeled "red_panda" appeared at the top of the Artificial Analysis Text-to-Image Arena leaderboard, ranked approximately 40 Elo points ahead of the second-place model. The leaderboard uses a crowdsourced voting method similar to chess Elo ranking: two models receive identical prompts and a human voter selects the better result. Because the submission was anonymous, several AI researchers and journalists speculated publicly about the creator before TechCrunch confirmed on October 28 that the model belonged to Recraft.
According to Dorogush, the codename came from actual users: during closed testing, people kept generating images of red pandas, which stuck as an informal identifier before the official V3 name was applied.
V3 was trained on Nebius AI infrastructure using a Kubernetes cluster managed via Kubeflow. The training process involved significant engineering challenges. Early runs operated at roughly one-eighth the expected speed due to network bottlenecks affecting GPU-to-GPU gradient synchronization. Recraft and Nebius engineers resolved the problem through GPU Direct RDMA configuration, patched NCCL libraries for collective communications, and a recompilation of PyTorch with static NCCL builds. These changes brought training speed to within approximately six times the original target rather than the full theoretical maximum, but the resulting model achieved state-of-the-art performance on the PartiPrompts benchmark, with over 50% preference against major competitors in head-to-head evaluations.
At the time of its release, Recraft V3 held an Elo score of approximately 1172 on the Artificial Analysis Text-to-Image Arena hosted on Hugging Face, with a reported win rate of 72%. This placed it ahead of:
| Model | Elo score | Win rate |
|---|---|---|
| Recraft V3 | 1172 | 72% |
| FLUX 1.1 Pro | 1143 | 68% |
| Ideogram v2 | 1102 | 63% |
| Midjourney v6.1 | 1093 | 64% |
| Stable Diffusion 3 Large Turbo | 1084 | 61% |
| DALL-E 3 HD | 984 | 51% |
Recraft V3 held the top position on the leaderboard for five consecutive months following its release. The Artificial Analysis arena uses crowdsourced pairwise comparisons rather than automated metrics, so the scores reflect human aesthetic preferences among the platform's primarily technical user base. The company itself published a comparative analysis noting that its model outperformed FLUX 1.1 Pro, Ideogram v2, and Midjourney v6.1 across categories including image-text consistency, visual quality, and instruction following.
One of the most practically significant distinctions of Recraft V3 is its ability to generate native scalable vector graphics directly from a text prompt. Most text-to-image models, including Midjourney, DALL-E, Stable Diffusion, and Flux (text-to-image model), produce raster images. Raster images store pixel data at a fixed resolution and become pixelated or blurry when scaled up.
SVG (Scalable Vector Graphics) files describe images as mathematical paths, shapes, and curves. Because the geometry is defined mathematically rather than as a grid of pixels, SVG images can be scaled to any size without quality loss. This property is essential for logos, icons, and other brand assets that must appear cleanly at both thumbnail and billboard scales.
Recraft V3 generates SVG files through a dedicated vector model variant. The outputs include structured layers and clean geometry. The vector capability spans a range of complexity, from simple monochrome icons and pictograms to detailed multi-color illustrations. Available vector styles include flat vector art, line art, engraving, bold stroke, and over 20 additional substyles.
The limitation is notable: generated SVG files frequently contain excessive anchor points and redundant path segments. Professional designers using V3's vector output for production work have reported spending anywhere from 30 minutes to over an hour cleaning up a single generated file in tools like Adobe Illustrator or Inkscape before it meets the standards required for client delivery. The Recraft team acknowledged this as an area of ongoing improvement, and Recraft V4 (released February 2026) emphasized cleaner SVG geometry as a primary upgrade.
Despite this caveat, no other major text-to-image model at the time of V3's release offered native SVG generation. Recraft V3's vector mode filled a gap that previously required either manual illustration, specialized design software, or post-processing of raster outputs through auto-tracing tools, which produce similarly noisy path data.
Text rendering is one of the historically weakest areas across AI image generation. Models like Midjourney, DALL-E, and Stable Diffusion regularly produced misspellings, garbled characters, and distorted letterforms even for short words or phrases. FLUX models improved noticeably on this in 2024 but remained inconsistent with sentences longer than a few words.
Recraft V3 introduced a substantially different approach. Recraft describes V3 as "the only model in the world that can generate images with long texts, as opposed to just one or a couple of words." The model can render full sentences and multi-line paragraphs within images, with legible typography that integrates visually with the surrounding composition rather than appearing as a separate overlay.
Beyond legibility, V3 adds positional control: users can specify where text should appear within an image, at what size, and with what approximate style treatment. This positional text placement was unique to Recraft V3 at the time of release; as of early 2026, the Recraft documentation still describes it as the only model with this capability.
Practical applications include:
The text rendering capability is available in both raster and vector output modes, though SVG text rendering is generally cleaner because vector paths describe letterforms precisely rather than approximating them as pixel grids.
Recraft V3 includes several features aimed at professional brand consistency workflows, which represent a departure from how consumer-oriented models like Midjourney and DALL-E operate.
The model ships with a style library containing over 100 curated styles organized into four main categories:
Realistic image styles are intended to produce outputs resembling photographs. Substyles include natural light, studio portrait, enterprise, HDR, hard flash, motion blur, black and white, evening light, faded nostalgia, retro realism, retro snapshot, urban drama, and product photo, among others.
Digital illustration styles cover drawn and computer-generated images with simplified or stylized textures. This category contains 46 distinct options including hand-drawn, grain, pixel art, pop art, risograph, clay, pencil sketch, bold sketch, street art, and urban sketching.
Vector illustration styles produce flat-color and limited-palette outputs appropriate for icons and scalable graphics. Sub-styles include line art, engraving, and bold stroke.
Emblem styles are designed for badge and logo-format outputs, with options including prestige emblem, stamp, punk graphic, and vintage emblem.
The default style for Recraft V3, used when no style is specified, is designated "Recraft V3 Raw." This mode gives the model more latitude in aesthetic decisions rather than constraining output to a specific visual preset.
Users can upload reference images to create a custom style profile. The model extracts the visual characteristics of the reference material and applies them consistently across subsequent generations. This workflow allows a brand team to establish a style template based on existing assets, photography, or illustration work, then generate new images that match that aesthetic without fine-tuning the model's weights. Style creation costs 40 API units ($0.04) per operation.
The API accepts exact hexadecimal color codes as parameters. Images generated with color specifications will use those colors without the color drift that typically occurs when prompts describe colors in natural language. This allows marketing and design teams to maintain brand palette compliance across batches of generated assets.
V3 exposes an "artistic level" parameter that adjusts how much creative latitude the model takes with composition and perspective. Lower values produce conventional, expected layouts; higher values result in more unusual angles, cropping, and compositional choices. This parameter gives users a dial between predictable commercial output and more experimental results.
Recraft V3 Raw is the model's default style, used when no explicit style preset is selected. Unlike named styles that constrain output toward a specific aesthetic category, Raw mode does not enforce a particular visual language. The model makes its own decisions about rendering approach based on the prompt.
In practice, Raw mode tends to produce results that lean toward photorealistic or highly detailed illustration output. It is described in Recraft's documentation as the most versatile option for users who want the model's full range without committing to a style category. For developers using the API, passing no style parameter defaults to V3 Raw behavior.
Raw mode is distinct from the model's photorealism substyles (such as natural light or studio portrait), which actively guide the model toward photographic rendering conventions. Raw mode leaves more decisions to the model's training distribution.
Recraft V4 was released on February 17, 2026, described by the company as a ground-up rebuild focused on what it calls "design taste." The model was developed in close collaboration with professional designers, with training optimized around aesthetic judgment rather than broad visual appeal.
V4 offers improved handling of composition, lighting, material surfaces, and spatial depth. The model follows complex positional instructions more reliably than V3, including foreground and background relationships, rule-of-thirds placement, and layered scene construction.
V4's text rendering is sharper than V3, with typography treated as a structural component of the composition rather than an overlay element. The SVG output quality is cleaner, with fewer redundant anchor points, addressing the main criticism of V3's vector files.
A new Exploration Mode generates multiple visual variations from a single prompt, returning a set of options to compare before committing to a refinement direction. Exploration Mode costs 16 credits per generation (2 credits per image across 8 images).
V4 is available in four variants:
| Version | Resolution | Generation time | Price (raster) | Price (vector) |
|---|---|---|---|---|
| V4 Standard | 1024x1024 | ~10 seconds | $0.04 | $0.08 |
| V4 Pro | 2048x2048 | ~28 seconds | $0.25 | $0.30 |
| V4 Vector | 1024x1024 | ~15 seconds | N/A | $0.08 |
| V4 Pro Vector | 2048x2048 | ~45 seconds | N/A | $0.30 |
V4 supports multiple output formats: SVG, PNG, JPG, PDF, TIFF, and Lottie (for animated assets). All V4 tiers are available on the Free plan. However, at V4's launch, some V3 features including style creation, prompt-based editing, image sets, and the artistic level parameter were not yet supported in V4.
The following table summarizes the key capability differences between Recraft V3 and major competing image generation models at the time of V3's release:
| Feature | Recraft V3 | Flux 1.1 Pro | Midjourney V6.1 | GPT Image 1 |
|---|---|---|---|---|
| Native SVG output | Yes | No | No | No |
| Long text in images | Yes | Partial (short phrases) | Limited | Partial |
| Text positioning control | Yes | No | No | No |
| Brand color specification | Yes (hex codes) | No | No | No |
| Custom style creation | Yes | No | Limited | No |
| Inpainting/outpainting | Yes | Limited | Yes | Yes |
| API access | Yes | Yes | Yes (alpha) | Yes |
| Raster output | Yes | Yes | Yes | Yes |
| Model size | 20B parameters | ~12B parameters | Undisclosed | Undisclosed |
| Arena ELO (Oct 2024) | 1172 | 1143 | 1093 | 984 (DALL-E 3) |
On aesthetic quality in head-to-head evaluation, Midjourney remains a top choice for artistic and illustrative styles, with particularly strong watercolor and painterly rendering. FLUX 1.1 Pro is generally faster (5-15 second median generation time vs. 7-10 seconds for V3) and is highly regarded for photorealism. GPT Image 1 (released by OpenAI in April 2025 as the successor to DALL-E 3) excels at instruction-following in complex compositional scenarios.
Recraft V3's differentiated position is in the combination of vector output, long text rendering, and brand control tooling, none of which are available together in competing models.
Recraft V3 is available through the Recraft web application, iOS and Android apps, and the Recraft API. Pricing follows two structures: subscription plans for application users and prepaid API unit packages for developers.
API access uses a credit system where $1.00 purchases 1,000 API units. Units do not expire. Per-operation costs are:
| Operation | Units | USD |
|---|---|---|
| V3 raster image | 40 | $0.040 |
| V3 vector image | 80 | $0.080 |
| V4 raster image | 40 | $0.040 |
| V4 Pro raster | 250 | $0.250 |
| V4 vector image | 80 | $0.080 |
| V4 Pro vector | 300 | $0.300 |
| Style creation | 40 | $0.040 |
| Background removal | 10 | $0.010 |
| Image vectorization | 10 | $0.010 |
| Creative upscale | 250 | $0.250 |
| Crisp upscale | 4 | $0.004 |
| Erase region | 2 | $0.002 |
The web application offers monthly and annual subscription tiers. Annual billing reduces cost by 20%. Plan tiers at the time of V3's release included:
| Plan | Monthly price | Monthly image credits |
|---|---|---|
| Free | $0 | 50 images per day |
| Basic | $10 | ~1,000 credits |
| Advanced | $27 | Higher allocation |
| Pro | $60 | ~8,400 images |
Images generated on the Free plan are owned by Recraft, displayed publicly in the community gallery, and do not carry commercial rights. Paid plans grant full ownership and commercial rights and allow private generation.
Subscription credits do not roll over at the end of each billing cycle. Additional credits can be purchased as top-ups. The API uses a separate prepaid unit system and is not tied to the subscription credit pool.
Recraft V3's combination of vector output and color control makes it one of the few AI tools usable early in a logo design process. A designer can generate vector illustrations in multiple styles with exact brand colors, export the SVG files, and begin manual refinement in vector editing software rather than starting from scratch. The workflow does not replace professional logo design, but compresses the ideation and draft generation phase.
The text-in-image capability addresses a longstanding pain point for marketing teams using AI image generation. Social media posts, ad banners, and email headers often require text overlay as part of the design rather than as a separate layer added in post-production. V3 generates the text as part of the initial image, reducing the number of production steps.
The flat vector and line art styles are suited to icon set production for web and mobile applications. The style library's consistency features allow a team to generate multiple icons in the same visual language without manual matching. The resulting files require cleanup but provide a faster starting point than purely manual work.
Recraft's canvas includes mockup generation tools that place generated images into product templates such as t-shirts, mugs, and packaging. These are useful for e-commerce sellers and product designers who need to visualize designs on physical objects without photography sessions.
The API's style and color controls make V3 suitable for automated batch generation workflows. A team can define a style template once and use the API to generate hundreds of on-brand assets programmatically for use across multiple campaigns or markets.
Recraft V3's October 2024 launch received substantial coverage after TechCrunch reported the anonymous "red_panda" model's appearance at the top of the Artificial Analysis leaderboard. The unusual route to public attention, appearing pseudonymously on a benchmark before the official announcement, created attention in AI research and design communities.
Joe Penna, a former researcher at Stability AI, publicly praised the model shortly after its reveal: "Wow! Amazing new model, Recraft. I'm very impressed."
Design professionals who tested V3 noted its performance on photorealistic detail. In head-to-head testing, photographer Jenn Mishra found that Recraft "produced more photographically complex images with detailed depth of field and realistic object placement compared to Midjourney." Across design forums and LinkedIn, the most frequently cited workflow benefit was the vector generation capability, with some designers estimating it saved two to four hours per project compared to manual illustration.
The model was noted by TechCrunch, Business Insider, and Decrypt as one of the more surprising releases of late 2024 given Recraft's low public profile relative to Midjourney and OpenAI. The Artificial Analysis leaderboard performance gave the company a credible benchmark anchor that larger, better-known models had not matched.
Recraft V3 remained on the top of the Artificial Analysis leaderboard for five consecutive months, a run that helped the company close its Series B in May 2025.
Enterprise adoption followed the Series B announcement. Recraft reported that Netflix, HubSpot, and Ogilvy were among organizations using the platform. The company reached 7 million users by early 2026, up from 4 million at the Series B.
The main practical limitation of Recraft V3's vector generation is path complexity. Generated SVG files frequently contain redundant anchor points and overlapping path segments that reflect the model's raster-trained underlying architecture rather than a true understanding of vector geometry. Professional designers report spending 30 minutes to over an hour cleaning up a single generated file before it meets production standards. This limits the time savings relative to the idealized workflow.
For scenes involving many distinct objects or precise spatial relationships between elements, V3 can produce inconsistencies in proportions and spatial arrangement. The same prompt may return noticeably different compositional structures across multiple generations, which reduces reliability for workflows that require consistent layouts.
V3's design philosophy prioritizes graphic design accuracy over photorealism as a primary goal. For projects that require true photorealistic output as the main objective rather than brand consistency or vector capability, models like FLUX 1.1 Pro or Midjourney may produce more convincing photographic results for a broader range of subjects.
Recraft experienced service reliability issues after V3's launch. Server capacity problems led to erroneous credit deductions for some users between September 2024 and January 2025. In February 2025, a service outage lasting over five days prevented users from accessing generated content; no advance notification was issued. Export and download failures were reported from February through April 2025. These operational issues drew criticism from users who depended on the platform for professional work.
Images generated on the Free plan are publicly visible and owned by Recraft rather than the user. This limits the platform's usefulness for commercial experimentation without a paid subscription, which is a different policy from competitors like FLUX (where models are available open-source) and Midjourney (where free-tier images are in a community gallery but the user retains rights under certain conditions).
The Recraft canvas application performs poorly on mobile devices and low-powered laptops, limiting access for users without desktop workstations or mid-range or higher computers.