Recraft

Recraft was founded in London in 2022 by Anna Veronika Dorogush, a Russian-born ML researcher who had previously led the development of CatBoost, Yandex's popular open-source gradient boosting library used widely across the machine learning industry. Recraft was her second major project, and she approached it with a very specific thesis that set the company apart from most other image generation startups: rather than building a general-purpose model that would compete with Midjourney on aesthetic quality or with Stable Diffusion on open-source reach, she decided Recraft would focus exclusively on the needs of professional designers. Not hobbyists, not social media creators, not marketers generating stock imagery, but the specific category of working designers who produce logos, branding, packaging, posters, and marketing collateral for commercial clients.

This narrower target market meant the product needed capabilities that general-purpose models ignored: vector output, precise text placement, brand style consistency, and the ability to generate full sentences of legible text inside images. All of these are table-stakes features for a designer and essentially missing from general-purpose image models. Recraft bet that a specialist model tuned for these capabilities would beat generalist models on professional design work even if it lost to them on raw aesthetic benchmarks. The bet paid off, by 2025 Recraft had more than three million users across over two hundred countries, a significant fraction of them paying professional designers rather than casual consumers, and the company was cited in the fal/a16z State of Generative Media report as the clearest example of how specialist models can win a vertical even while general-purpose models dominate headlines.

Recraft's architectural approach has not been fully disclosed but the technical capabilities suggest it combines a standard diffusion transformerTransformerThe neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks. backbone with several specialized modules trained on top. The vector output capability is implemented by training the model to generate SVG code as an alternative output format, which is technically unusual, most image models generate raster pixel grids and Recraft treats vector graphics as a first-class output mode. The text rendering capability is built on training data that includes enormous volumes of images with paired precise text transcriptions, which let the model learn character-level typography rather than just word-level. The brand style consistency feature uses style conditioningConditioningAny extra input that steers generation beyond the prompt, such as a depth map, pose, edge map, or reference image. at inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. time (passing reference images through the encoder to extract style features that get merged into the generation pipeline) rather than fine-tuningFine-tuningContinuing to train an existing model on new data to specialize it for a style, subject, or domain, without training from scratch., which is why it works instantly without any setup or training.

The practical consequence of these architectural choices is that Recraft feels like a different kind of tool than the other models in this section. When you use FLUX or Midjourney you write a promptPromptThe text description you provide to a model to specify what you want it to generate. and accept whatever the model produces. When you use Recraft you get a design interface with controls for typography, positioning, style references, vector versus raster output, and the model generates within those constraints. The prompt is one input among several. This is much closer to how working designers actually think about the tools they use, and it is why Recraft has been able to sell to design teams that other image models cannot reach.

Recraft V3 (October 2024)

Released under the internal codename 'red_panda.' For a small specialist lab to hold its own against every better-funded generalist for months was the moment the field stopped treating Recraft as a design-tool curiosity. Through late 2024 and into 2025 its V3 traded at the front of the pack with Midjourney, OpenAI, Stable Diffusion, and FLUX 1.1 [pro]. What made that possible is the more interesting part. V3 was the first model that could reliably render full paragraphs of legible text inside images, which alone would have made it notable, and it shipped the vector output and precise text positioning features that are still Recraft's core differentiators.

Recraft V4 Pro (early 2026)

The current flagship as of mid-2026. V4 Pro brings significantly improved photorealistic quality (closing most of the gap to FLUX.2 and Nano Banana Pro for design-oriented photography), better handling of multi-element compositions, and expanded vector output capabilities including complex illustrations and icon sets. The brand consistency features have been enhanced with a 'brand kit' system that lets teams upload logos, color palettes, and typography specifications and have the model respect them across all generations. Pricing is competitive with Ideogram and FLUX at the professional tier.

Recraft's strengths are exactly the things designers need: legible long-form text, vector output, precise text placement, brand consistency, and a design-first interface. If you are producing posters, logos, packaging, marketing collateral, social media ad variants, or any structured design output where professional polish matters, Recraft is typically the best single model to use. The vector output in particular is unique among the major labs and makes Recraft the only option if you need scalable graphics that can be edited in Illustrator or Figma.

Recraft's weaknesses are photorealism and aesthetic range. It is good at design-polished photography but it is not the best model for hero photorealistic assets (FLUX.2 and Nano Banana Pro are better) and it is noticeably weaker on stylized or artistic work (Midjourney 7 is much better). The model is also entirely closed-source, meaning no LoRAs, no fine-tuningFine-tuningContinuing to train an existing model on new data to specialize it for a style, subject, or domain, without training from scratch., no open weightsOpen weightsModels whose trained parameters are publicly downloadable, allowing anyone to run them locally or fine-tune them. Contrasts with closed/API-only models., and no offline use. For the professional designer market this trade-off is fine because the alternative models in that market are also closed; for the open-source community Recraft is effectively invisible.

Recraft occupies an unusual strategic position as one of the only major image labs that has successfully specialized in a specific vertical and won that vertical. Most other labs are competing on the same general-purpose axes (photorealism, promptPromptThe text description you provide to a model to specify what you want it to generate. following, aesthetic quality) and trying to beat each other on benchmarks. Recraft picked a different game entirely: own professional design work, make features specifically for working designers, and be invisible to the consumer hype cycle. This has made Recraft structurally immune to a lot of the competitive dynamics that buffet the other labs. When GPT-Image-2 launches and the text-rendering benchmarks shuffle, Recraft does not lose its customers because Recraft's customers are not buying text rendering, they are buying a design tool that happens to use an image model. The fal/a16z report's finding that operators need many specialized models rather than one general-purpose model is essentially a description of the world Recraft has been building for since 2022.

Recraft is the odd one out in this chapter because it is aimed at professional designers, not general image-making. It shocked the field in late 2024 when a model that appeared anonymously on the Artificial Analysis arenaArenaA public leaderboard that ranks models from large numbers of blind pairwise human preference votes, which tends to reward average appeal over fitness for a specific job. under the codename red panda turned out to be Recraft V3 and topped the board, beating Midjourney, FLUX 1.1 Pro, and DALL-E 3. Its differentiators are real design deliverables: long, correctly spelled in-image text, and, uniquely, generating genuinely editable vector SVG files from a promptPromptThe text description you provide to a model to specify what you want it to generate., plus brand colors, reusable custom styles, and an infinite-canvas editor. You work in it through the web app or the APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface., also on fal and Replicate, on a credit-based freemium plan.

So the way to get value is to use it for what general models do badly: logos, icons, and illustrations as editable vectors, on-brand asset sets with locked styles, and images that must carry accurate legible text. It is worst at cheap high-volume bulk generation and at free-form artistic exploration, where its design-first tuning and credit costs work against you. And be honest about the leaderboard: the late-2024 number-one position has since been overtaken, so choose Recraft for the vector and brand workflow, not because it is still the single highest-scoring image model.

Recraft is iterating fast on its own model family, with V4 in February 2026 rebuilt around what it calls design taste, meaning deliberate choices about composition, color, lighting, and material realism rather than literal promptPromptThe text description you provide to a model to specify what you want it to generate. matching, and V4.1 a few months later sharpening photorealism, illustration, and cleaner vector defaults. The strategic bet, backed by Accel, Khosla, and Madrona, is that professional design is a distinct, underserved vertical that general models handle poorly, and that owning editable vectors, dependable typography, and brand consistency is more defensible than competing head-on with OpenAI, Google, and Midjourney on general image quality. Founder Anna Veronika Dorogush, who built the open-source CatBoost library before Recraft, frames it as being excellent at what is mission critical, which is why the company trains its own models rather than wrapping someone else's.