Recraft: The Image Has to Work

The Panda and the Deadline

On October 30, 2024, Recraft announced V3, an image model that had been tested under the name Red Panda. It had reached the top of the image preference leaderboard run by Artificial Analysis. A small company's model was suddenly being discussed alongside the biggest names in artificial intelligence. That was a good way to get noticed. It was also the beginning of a story that becomes much more useful when we stop treating the leaderboard as the destination.

An image can win a comparison and still make a terrible employee. It can be beautiful, convincing, and completely unable to give you the matching illustration you need next Tuesday. It can draw an impressive poster whose date takes longer to repair than the poster took to generate. It can produce a handsome logo which becomes an archaeological excavation when you try to change one curve. The first look earns attention. The work that follows decides whether a tool belongs in a business.

That is the test I want to apply to Recraft, Pär. You already know how entertaining an image model can be, and you have spent enough time around local models, rented graphics processors, and ComfyUI to recognise the difference between a demonstration and a useful workflow. Another attractive gallery is an invitation to experiment. It is a weak argument for changing the way you work. The interesting question is which particular jobs become easier with Recraft, and which parts of your existing approach remain valuable.

This is a researched assessment as of October 6, 2026. I have checked current documentation, announcements, pricing, and terms. I have not run a controlled image comparison, and I will describe proposed tests as proposed tests. That distinction matters here because much of the public discussion mixes results from different generations, combines the company's models with other models inside its application, and repeats marketing claims after the evidence behind them has aged.

The episode has a working proposition. Recraft deserves attention wherever you need a visual language that can survive repeated use: related illustrations, editable artwork, or a collection of small graphics that clearly belong together. It also deserves a look when you want to explore composition quickly. Whether it earns a place depends on the exported files and the time required to finish them. We will follow that proposition from the company and the technology into the practical business of using an asset.

There is a reason that practical end matters especially in your world. A one-person operation cannot hide ten minutes of repair inside somebody else's department. The person choosing the illustration may also be writing the article, laying out the page, answering the booking enquiry, or fixing the software. A tool that reduces those demands is interesting even if it never wins an image beauty contest.

By the time we reach the end, there will be a concrete experiment: a small collection of real briefs, a few candidate images for each, an actual output test, and a way to count the cost that includes your attention. There is also a contractual detail that could disqualify Recraft from one of your more tempting uses before the first image is generated. Its commercial permissions and its model-training permissions are different. We will get to that after establishing what the company is trying to build.

The Company Behind the Canvas

Recraft is attempting a difficult double act: building the engine that generates images, and the workspace in which people decide what those images should become. Those businesses demand different skills. A model must learn useful visual behaviour. An application must make that behaviour accessible, repeatable, and worth returning to. Owning both gives Recraft opportunities that an ordinary software wrapper would struggle to match, along with considerably more work.

The company says it was founded in the United States in 2022 and is headquartered in London. Its UK company, Recraft AI Limited, was incorporated that December. The current developer agreement identifies Recraft Inc. as the contracting party and gives a San Francisco mailing address. That combination explains why descriptions of the business sometimes appear to disagree about its geography. Headquarters, incorporation, and the company on your contract answer different questions.

Its founder and chief executive is Anna Veronika Dorogush. Before Recraft, she led machine-learning tools work at Yandex and helped create CatBoost, having also worked at Google, Microsoft, and ABBYY. Her name appears on the 2018 CatBoost paper alongside Vasily Ershov and Andrey Gulin. That is a substantial technical history behind the founder biography. She had already spent years making machine learning useful to people who needed dependable software.

CatBoost deserves a brief explanation because its name can sound suspiciously appropriate for an image company. It is a gradient-boosting library. Broadly, that means combining decision trees so that successive stages improve the predictions of the overall system, including when the data contains categories. Think structured prediction problems rather than drawing cats. CatBoost establishes relevant experience in machine-learning engineering; it tells us nothing by itself about the architecture inside Recraft's image models.

The funding history shows how deliberately Recraft pursued its own technology. In January 2024, it announced a twelve million dollar Series A led by Khosla Ventures and Nat Friedman. The announcement connected that investment to development of its own foundation model and controls for professional design. Consistent styles and related visual assets were already part of the ambition. Those priorities preceded the later attention surrounding its performance against other image generators.

In May 2025 came a thirty million dollar Series B led by Accel, with Madrona and existing investors participating. A July extension then brought in NVIDIA and design executive Ethan Eismann. The size of that extension was not disclosed. The careful arithmetic is forty-two million dollars across the announced Series A and initial Series B, followed by an unspecified extension. That establishes backing; it does not establish profitability, financial security, or how much cash remains today.

What were those investors buying into? Their published arguments emphasised the relationship between capable models and the practical demands of designers. My reading is that Recraft wants to compete over the distance between an attractive generation and something a professional can confidently use. Image quality attracts attention. Control determines how much work follows. If a customer regularly spends more time repairing the output than generating it, a faster generator only solves part of the problem.

This also explains why the company, the model family, and Recraft Studio must remain distinct in our heads. Recraft trains proprietary models. Recraft Studio is its application, and today that application also offers models from other providers. The Recraft application programming interface, or API, gives developers access to its own capabilities from their software. Choosing a model inside Studio and calling a Recraft model from your own code are different product decisions, even when the resulting image looks identical.

That distinction can change the credit you give the model. Studio offers natural-language editing through outside providers, including Google and Qwen, and an automatic mode can choose the engine. If an edit preserves a face well, check which model made it. A successful edit inside Recraft Studio establishes something about that chosen workflow; it does not automatically establish the same ability in Recraft’s own image model.

There is an interesting strategic tension in offering competitors' models beside your own. My interpretation is that it lets Recraft compete for the working environment even when another model is better suited to a particular request. The customer can remain in the application while changing the engine. Meanwhile, the company's own models give people a reason to choose Recraft specifically. That is a plausible strategy; public announcements do not tell us whether it produces stronger retention or better margins.

The difficult part is paying attention to both sides. Research asks which experiments could improve the model. Product development asks which frustrating interaction should disappear this week. Owning the model can let a company respond to requirements that a supplier would otherwise control. But it also creates a demanding research programme alongside the everyday work of running software. Funding provides room to attempt that combination. Execution decides whether the combination becomes an advantage.

For you, Pär, the useful conclusion is a question of boundaries. Evaluate the lab for what its models produce, the application for how well it supports your decisions, and the API for how comfortably it fits software you control. Those judgments can produce different answers without contradiction. You can value a company's research without adopting its entire workspace. Recraft has placed a substantial bet on owning both generation and the creative process; you get to decide how much of that bet belongs inside your operation.

Teaching Pictures to Read

A hypothetical poster for a small event makes the problem clear. The picture is excellent, the atmosphere is right, and the lettering confidently announces a date that does not exist. You have received something visually convincing and commercially unusable. That little disaster explains why text matters so much to a company pitching image generation to designers. A picture can survive a slightly unusual cloud. A telephone number has considerably less room for artistic interpretation.

Letters have an awkward double life in an image. They are shapes, with spacing, weight, curves, and texture, but they also carry an exact symbolic sequence. Change the curve of a painted hill and you may improve the composition. Change 1 digit in an address and the parcel goes somewhere else. Design work constantly crosses that boundary between visual judgment and exact information, which makes it a particularly revealing test of image generation.

Recraft offered a useful account of its approach in November 2024, describing V3. Its team trained an optical character recognition system, or OCR, to identify text and positions, then used those layouts to train a captioner that included the actual words. A large language model generated text layouts from prompts. The image generator received a rendered layout alongside its caption, using a conditioning approach the team compared with ControlNet. Layouts could also be specified manually. The important historical detail is that the company described several cooperating components, with an explicit representation of where words belonged.

That account makes the capability easier to understand. Planning where the text goes gives the image generator a more specific instruction than an ordinary description of a scene. The product can expose that planning stage to its user. It also shows why a feature such as text positioning deserves attention independently of how attractive the finished picture looks.

For a poster, I would separate the decisions before asking any model to help. What needs to catch the eye, what must be read immediately, and what must remain exact? A concert title might benefit from wonderfully strange lettering. A booking deadline benefits from being correct, legible, and easy to change. Those are different jobs even when they appear in the same rectangle, and they deserve different levels of creative freedom.

Consider a promotional graphic for PartyPär. You might welcome an exuberant illustrated popcorn machine, a ridiculous cloud of popcorn, and colors that make a beige accountant briefly reconsider their life. The machine illustration is an opportunity for interpretation. The rental price and booking details are specifications. My recommendation would be to generate the visual ingredient, approve it, and place the exact copy with ordinary typography so the next revision remains a small editing task.

The same distinction matters when discussing what we know about the model itself. The V3 account describes a particular historical implementation. It gives us no basis to announce that V4.1 has the same architecture, or to invent a current parameter count. In the current material reviewed for this episode, I found no published downloadable weights or supported local inference route. We can describe the documented behavior with confidence while leaving the undisclosed machinery undisclosed.

Style introduces another version of the same problem. Suppose 6 illustrations share thick rounded lines, restrained colors, and an affectionate sense of absurdity. You could reasonably call that a consistent style even if the blue shifts between images. A brand specification might require the blue to be identical. Similarity is something you judge across pictures; an exact fill value is something you can set and inspect in a file.

I would therefore evaluate a design model with 2 questions. Did it make a visual decision that I actually wanted? And can I turn the result into an asset whose important details I control? For Swedish material, that includes checking every accent and every letter, especially when lettering becomes decorative. The entertaining part of image generation is watching a suggestion become visible. The useful part is deciding which decisions to keep, which to correct, and which should stay in the hands of ordinary design tools.

Choosing the Actual Model

Now we need to identify which Recraft we are talking about. V3 arrived in October 2024, V4 followed on February 17, 2026, and V4.1 launched on May 14. A dedicated V4 Styles family arrived in August. V4.1 Flash followed on September 23. An old review may therefore describe a real capability accurately while recommending the wrong model for the job you have today.

The useful way through the names is to separate the choices. Main V4.1 is the expressive option, while Utility aims for simpler scenes and more predictable composition. Raster gives you pixels; Vector gives you editable Scalable Vector Graphics files, usually called SVGs. Pro adds detail, with the raster version generating at approximately 2,000 pixels on a square side instead of 1,000. Utility, Vector, and Pro can be combined, so these names describe intersecting choices rather than a single ladder from bad to good.

Consider a picture of a coffee cup. An expressive interpretation might choose dramatic lighting, an unusual crop, and a surface that suggests an expensive magazine. For an early campaign idea, that could be exactly the contribution you wanted. For a catalog layout, you may want the cup facing forward with room around it and nothing competing for attention. I would choose between the main and Utility directions according to that brief, then judge the output against it.

The same logic applies to paying for more detail. If the composition is wrong, a more detailed version gives you a more detailed problem. If you are still choosing between 12 ideas, cheaper exploration may be useful. Once a direction earns its place, higher quality becomes easier to justify. The extra expense earns its place when it improves a direction that already works.

There is also a slightly untidy feature boundary. The current API documentation supports image-to-image with V4.1, while V3 still handles the documented inpainting, outpainting, and background replacement operations. V3 also retains explicit text placement and negative prompts. V4.1 supports preferred colors and custom styles, but lacks those particular controls. A higher version number does not automatically inherit every button from an earlier model, so the operation matters as much as the name.

V4 Styles deserves its own choice. It is a dedicated family for generating in a supplied visual style, and requires a reference or a saved style. Its references provide a reusable direction across different subjects. If the assignment is a coordinated series, that is a compelling capability to investigate. If the assignment is to preserve the exact identity of a particular person or product, I would evaluate that separately instead of treating a style match as proof.

Flash serves another purpose: fast raster exploration. Its API has no vector version, no Pro variant, no style references, and no image-to-image support. Those limits matter because rapid variation and reference-driven consistency solve different problems. A sensible candidate workflow is to explore an idea quickly, select a direction, and use Refine Details on the chosen image. Refine Details uses V4.1 Pro to redraw small details rather than merely sharpening them, so faces, lettering and product labels need another inspection afterward.

Recraft's published speed comparison used 100 prompts, with 1 run each, measured from Kansas on September 21. Flash returned a 1.4-second median, tied with FLUX.2 klein 9B, and a 1.8-second 95th percentile against klein's 2.1. Recraft used its own API; competitors ran through fal or Runware, and download time was excluded. The headline 1.3-second figure came from a separate, larger prompt set. This supports a claim of fast performance in that test, while leaving global speed supremacy unproven.

There is a useful reality check on quality, too. On the live Artificial Analysis overall text-to-image board checked for this episode, V4.1 Utility ranked 25 and Utility Pro ranked 31. The ordinary V4.1 ranked 51. These are human preference results for generated images, with uncertainty around the positions. They do not score SVG editability or the time required to finish a branded illustration. Recraft's historical leaderboard crowns belong to their dates, and they should stay there.

Those results suggest a useful next test: a coordinated set of graphics and an awkward brief that demands restraint. Taste sometimes means resisting the opportunity to decorate. Recraft should get a chance to demonstrate its particular design strengths. A series of related assets gives us a better place to examine those strengths than another isolated gallery image.

The Seventh Icon

The revealing test for a design model is often the 7th icon. Suppose the popcorn looks wonderful, the speaker looks wonderful, and the bouncy castle looks wonderful, but they appear to have been commissioned by 3 companies that have never met. Each picture passes its individual audition, while the collection refuses to become a cast. The task has changed from making an attractive object to establishing rules that survive a change of subject. That is where Recraft becomes especially interesting, because a useful visual style needs to travel beyond the image that introduced it.

Recraft’s V4 Styles system accepts 1 to 10 reference images and lets you save the resulting style for later generations. It offers precise and flexible matching, giving you a choice about how closely the reference should guide the result. You can supply a new subject while retaining that saved visual direction. That makes an approved illustration a possible starting point for a larger body of work. The attraction is the shorter journey from having something you like to having a collection that looks as though somebody deliberately designed it together.

The choice of reference deserves more attention than the number of references. Recraft’s own guidance recommends a strong image or a set that already agrees with itself, because unrelated material can introduce competing motifs and visual treatments. Supply a restrained newspaper drawing alongside a glossy party advertisement and a photograph of a forest. You may understand the common thread in your head, while the model has been handed several different instructions at once. A useful reference set makes its visual rules visible, leaving the prompt to explain what should happen within them.

Choose Recraft’s vector generation and the deliverable can be an actual SVG with editable shapes. Recraft also provides tools for adjusting individual points, curves and colours, so there is a practical route into the result after generation. Picture a popcorn illustration whose overall silhouette works, but whose container needs a different colour and whose bottom edge feels untidy. Those become potentially manageable edits to an asset you already want to keep. The value lies partly in being able to finish the picture without asking the model to reinterpret every good decision it has already made.

Recraft also has competition in vector generation. Adobe’s Firefly documentation describes generating vectors from text and downloading SVG files. That establishes another option in the category, without telling us which one will handle your particular brief better. For one conventional icon, an existing set or an ordinary drawing tool may be quicker still. The distinctive case for Recraft is worth testing against those alternatives, particularly when you need a collection that grows over time.

There is a distinction here between a convincing brand palette and an exact colour specification. Recraft explicitly warns that matching the reference colours is not guaranteed, which makes checking them part of the work. PartyPär’s pink should ultimately be a deliberate production choice, with the appropriate colour values applied to the finished artwork. You might get an illustration that captures the right energy but arrives with a slightly different pink in each of its main shapes. Editable fills give you a way to bring those shapes into agreement, and the generated image becomes a starting point for controlled finishing.

Style consistency also deserves its own boundary around identity. A collection can share beautiful lighting, brushwork and colour while quietly changing the person who appears in it. For an imaginary character, some variation may be acceptable; for a recognisable person in a business campaign, the face carries a different responsibility. A man wearing a bright pink suit might communicate the idea of PartyPär without being a convincing image of Pär. The same problem applies to objects, because an illustration of a speaker can fit the collection while depicting a speaker you do not actually rent out.

Recraft’s character guidance says it has no dedicated character-tracking feature, and that should shape expectations. I would assess resemblance and visual treatment separately whenever the identity of the subject matters. Try reviewing a sequence by temporarily ignoring the backgrounds and comparing only the faces, then reversing that attention and examining the collection as a whole. Those are different questions, and success on either question can conceal failure on the other. A saved style gives you something useful to hold steady, while accurate people and products still require their own scrutiny.

The asset I would want from this process has room for ordinary human decisions. It should tolerate a different crop, a corrected fill, a smaller placement and the arrival of another illustration beside it. That shifts attention away from whether an image wins a moment of admiration on a generation screen. The more revealing question is whether you can comfortably live with it through the rest of the job. When the 7th icon belongs with the others and you can still edit the whole collection sensibly, the model has contributed something more durable than an attractive preview.

What Survives the Page

The printed page is a useful place to become less impressed by previews. A picture that fills a screen has plenty of room to announce its texture, while a small illustration beside a column has to communicate quickly. Fine details can become visual clutter, and a graceful thin line can lose its authority when the whole object shrinks. Judge a design while it is still surrounded by the actual headline, body text, neighbouring advertisement and white space. That setting reveals whether the image belongs to the publication, rather than merely looking attractive in isolation.

The arithmetic behind resolution is straightforward enough to keep in your head. At 300 pixels per inch, an image 1,024 pixels wide covers about 8.7 centimetres, and 2,048 pixels covers about 17.3 centimetres. Those figures describe how widely the available pixels are spread across the paper. They do not tell you whether the illustration is well drawn or whether its smallest lettering is readable. An image can have enough pixels and still be badly suited to its placement, just as a simple, well-designed graphic can survive conditions that would make delicate detail disappear.

An A4 page at 300 pixels per inch requires approximately 2,480 by 3,508 pixels before allowing for bleed. That is why a 2,048-pixel square should not automatically be treated as a full-page image at that resolution. Changing a resolution setting rearranges the relationship between pixels and physical size; adding more pixels involves resampling or another enlargement process. Neither operation removes the need to inspect the finished result at the intended size. The actual production requirement comes from the publication and its printer, so the useful question is how the asset will be placed and reproduced.

Vector output changes some of that arithmetic, but it introduces another set of questions. SVG can contain embedded raster images, so the file extension by itself does not establish that every visible element is a scalable shape. Even when the artwork consists of paths, useful editability depends on how those paths are organised and how much repair they need. Suppose you open a promising illustration and discover that the detail you want to adjust consists of dozens of overlapping fragments. I would judge that file by the ease of the necessary edit, alongside how it looks on the page.

Typography deserves a particularly Swedish examination. The words Årebladet, Mörsil and Undersåker give the model more to get right than a short English slogan, and a missing mark can be easy to overlook while admiring the composition. Add a date, a telephone number or a booking instruction and the consequences of an error become practical. I would read each piece of generated lettering as text, then inspect its spacing and shape, then look at it at publication size. That sequence helps prevent an appealing illustration from persuading you that you have already checked what it says.

Recraft supports manually placed text and uploaded fonts, which provides a sensible way to keep ordinary copy under direct control. A supplied font still needs the glyphs required for the language being used. Generated lettering has an enjoyable role when its physical appearance is part of the idea, such as a PartyPär title apparently assembled from popcorn. The booking details beneath that title can remain ordinary, editable type. That arrangement lets the imaginative part of the picture be imaginative while giving the information a clear and efficient correction path.

Factual graphics need a similar division of responsibilities. Take an illustration accompanying an explanation of the municipal budget: a school, a road and a meeting room can help introduce the subject. The amounts, the relationships between them and the lengths of any bars should follow the actual data through a controlled rendering process. A map needs that same discipline around locations, boundaries and labels, even when a decorative mountain sits beside it. An incorrect map in an infinitely scalable format is an infinitely scalable mistake, and clean curves do not make its geography any more trustworthy.

The last stage belongs to the whole document. Colour handling, page dimensions, bleed and Portable Document Format, or PDF, export settings need to suit the real print job, while the image should be checked after it passes through that workflow. I would place a candidate in a representative page early enough that an awkward crop or weak line can still influence the design. That also makes the evaluation feel like ordinary publishing work instead of an endless audition of isolated images. The useful asset is the one whose visual idea remains clear after all the surrounding decisions have been made.

Where I Would Use It

For Årebladet, I can see a useful role in recurring illustrations that give familiar civic subjects a recognisable visual treatment. A school, a newspaper arriving at a household, a road being repaired and a room prepared for a public meeting could belong to the same illustrated world. Consistency across those subjects would have a practical editorial value. Readers would encounter a publication making deliberate choices across different subjects, even when the available space changes. A small collection that can be adapted carefully may contribute more than a spectacular image with only a single possible use.

Local journalism also gives you a strong reason to keep the purpose of each image clear. A symbolic illustration of a meeting can accompany an explanation, while a picture presented as evidence of an actual meeting has to answer different questions. The location, the people and what happened there matter because readers may know all of them. I would give generated artwork a visible illustrative role and retain the actual documentary material wherever that is what the story needs. That leaves room for expressive design while respecting the particular value of a newspaper reporting on places its readers recognise.

For PartyPär, I would look at coordinated campaign artwork and supporting illustrations around the actual equipment. Popcorn, lighting and music offer plenty of visual material without requiring every image to become a literal catalogue photograph. Consider a flyer and a social image sharing the same drawn shapes, colour decisions and sense of movement, with the product photograph and booking copy placed deliberately inside the design. That is an appealing use of a generated visual language. Whenever a customer needs to understand exactly what they are hiring, the real object remains the reference that the finished material has to represent accurately.

For Pärception, the interesting opportunity is to give technical ideas a consistent and understandable visual character. A small illustration might show work arriving in a queue, a person making an editorial decision, or a process returning an item for correction. Those examples have enough meaning to guide the artwork before any choice of texture or lighting enters the conversation. Precise system diagrams can still carry their actual components and relationships, with illustrations supplying context or a memorable metaphor alongside them. I would want the pictures to help somebody understand the work, so each image needs a clear job before it needs a style.

For something like Tidningsgrafik.se, if you choose to pursue that direction, the architecture I would want is a bank of approved assets used by your own templates. The approach you have described around web templates and structured data already provides a place to assemble the factual page. Recraft could contribute a finished illustration or motif that gets approved before any customer orders a version. After that, a municipality name or updated figure can change through the template without asking an image model to draw the surrounding decoration again. That is the desired behaviour; whether Recraft’s standard terms permit that particular bank is a separate question, which we will come to shortly.

For assets whose licensing permits that archive, I would keep the approved files alongside enough information to understand where they came from. That means retaining the references, prompt, model identifier and saved style identifier, together with the final editable asset and a useful preview. The image links returned by Recraft’s API are temporary, lasting approximately 24 hours, so downloading the results belongs in the creation process. The archive should also make your finishing decisions visible, because the corrected asset may be more valuable than the original output. A future change then begins with something you can inspect and edit, instead of a memory of a promising generation.

Suppose a customer needs the same page for another municipality late at night. The figures change, the map is rendered from the appropriate geographical data, and the local names enter the text fields. The already approved illustration can simply be placed where the design expects it to appear. There is no creative uncertainty around that part of the order, because its job was settled earlier. New artwork can still be commissioned when the subject calls for it, while routine variations use the work you have already brought to a publishable state.

That arrangement fits a way of working in which you keep a clear boundary around each external service. Recraft can supply an asset, and your own files and rendering process can carry it through everything that follows. Your existing image workflows remain available for jobs where their particular controls are useful, including the personal references and trained assets you have already invested time in understanding. I would give Recraft work that exposes its strengths and judge the result after the ordinary finishing has happened. Its place in the toolkit would grow from the jobs it helps you complete comfortably and repeatedly.

The Price of Control

Recraft's current free tier comes with a restriction that changes the evaluation. The pricing page says free generations are public and are not licensed for commercial use. Paid generations have commercial rights and remain private, including after cancellation. A free account can let you investigate the interface and the sort of images it produces. When the experiment becomes a possible Årebladet illustration, PartyPär advertisement, or asset for a customer, the applicable generation rights become part of the choice.

The sharper detail is in the August 2026 terms. The paid Studio provision assigns whatever copyright Recraft has in the generated assets to the user, but also prohibits using those assets to train artificial intelligence models. The developer agreement applies a corresponding restriction to API outputs. Paying through that route does not turn generated images into unrestricted material for training a different model.

For someone who experiments with LoRAs, that is consequential. A LoRA is a low-rank adaptation, a compact trained addition used to change a model's behaviour. One tempting plan would be to generate a beautiful reference family in Recraft and then train a local style adaptation from those outputs. Under the current stated restrictions, that is not an authorised use. Your own original reference artwork is a different asset from the images the service produces. Keep that distinction clear when deciding what belongs in a training dataset.

You can see the business logic without having to like every part of it. A company selling access to a proprietary model has an incentive to prevent its output becoming a cheap route to reproducing that model's behaviour elsewhere. That is my interpretation of the restriction, not a claim about an internal company discussion. For a user who values portability, the effect is more immediate: ownership of an exported file and freedom to use that file for every future purpose are separate questions.

There is a second distinction in the data flow. Recraft's July 2026 training disclosure says individual-account content is used for training by default, with an opt-out in settings. Teams and API content are excluded by default. The API agreement also says Recraft does not train its models on API-generated assets, while third-party services can have their own terms. A private gallery setting therefore answers a different question from a training setting. Choosing another provider's model inside Studio adds another provider to consider.

The same disclosure describes billions of image and text examples from several sources: public repositories, third parties, users, and internally produced or synthetic material. It identifies licensed and public-domain material alongside copyright-protected material used under the company's stated fair-use position. That does not amount to an exclusively licensed training collection. Nor does the disclosure by itself settle the legal status of every training example. It gives us a more specific account than the vague phrase proprietary data would provide.

There is one more provision directly relevant to the illustration bank we discussed. Section 3.6 of the developer terms limits local caching to 30 days for application performance, and prohibits persistent content databases or repositories that bypass end users calling the API. That sits awkwardly beside the output-ownership grant. I would get written clarification before making Recraft’s API the source of a permanent reusable bank for an automated product. I am not interpreting it as a blanket ban on keeping an ordinary finished design; I am saying that this specific architecture needs clearer permission.

Those terms do not make the product unusable. They help define the job it can do for you. A commercially permitted illustration can become a reusable ingredient in a page or an advertisement. Your own source data, layouts, references, and approved outputs can stay under your control. The model that makes new images remains a service you depend on, and the permitted uses of its outputs remain bounded by the agreement. That is a manageable relationship when it is a deliberate choice.

I would also separate buying an image from buying certainty about the subject it depicts. A permission from Recraft does not tell you that a generated wordmark is distinct from every existing brand, or that an invented product detail matches something you actually rent out. Those questions still require ordinary professional judgment. The useful habit is to decide where the image is allowed to improvise before you ask for it. Then judge the result against that boundary, along with the commercial terms.

An Experiment Worth Finishing

The current API prices make a focused experiment inexpensive enough to be interesting. Flash costs seven tenths of a US cent per image. Ordinary V4.1 raster costs three and a half cents, and its standard vector output costs eight cents. The dedicated V4 Styles line has a five-cent standard vector option. Its Pro vector option costs twelve cents, compared with thirty cents for V4.1 Pro vector. Those are different model choices, so the word Pro alone does not identify either the bill or the behaviour.

Recraft's API uses prepaid units, bought separately from Studio subscription credits. The published pricing says purchased API units do not expire. For an occasional user, that separation may matter more than a subscription's impressive monthly generation allowance. You can evaluate a small callable service without committing your whole creative process to a new workspace. Whether Studio is worth paying for separately depends on how much you value doing selection and editing inside its canvas.

Here is the experiment I would actually propose. Give it an hour initially, with permission for the hour to end in a rejection. Choose two references you have the right to use and ten briefs that represent something you might publish. One reference could define a family of small icons, the other a family of larger illustrations. Make the subjects different enough to expose whether the style holds together: a machine, a person, a building, an object in motion, and a scene with several elements.

Produce two candidates per brief using the standard Styles vector model. Those 20 images consume one dollar at the current rate. Creating two reusable styles adds one cent. Make 20 Flash concepts for the separate question of rapid visual exploration, consuming fourteen cents, and use Refine Details on two selected raster results at twenty-one cents each. The total for those specified Recraft operations is one dollar and fifty-seven cents. That is consumed usage, before tax, extra attempts, competing tools, or any minimum amount you must purchase.

Run representative matching briefs through your existing workflow as well. If Qwen or FLUX supplies a raster result where you need a vector, count whatever conversion and editing the comparison actually requires. If your own workflow is much better at preserving a person or a product, give it a chance to demonstrate that too. A fair comparison keeps the brief and the intended deliverable stable. It does not require every tool to be used through an identical interface or an unsuitable prompt.

Then make the first decision with the model names hidden, if that is practical. Select the images that satisfy the brief, not the images you most want the experiment to vindicate. After that, unhide the names and record the finishing time. Recolour the shapes. Fix the spacing. Place the graphic on a representative page. Export it through the process that would actually deliver it to a reader or customer. The winner must survive that part of the afternoon.

The measurement is straightforward: generation charges, any rented graphics processor cost, and a value for the human time, divided by the number of accepted assets. For a family of icons, also count the cost of an accepted coherent set. Eight individually attractive images may still leave you with no usable system. And a cheap image that demands repeated repair can cost more than a dearer one that is ready quickly. You choose the hourly value; the experiment should reveal the minutes either way.

Keep the speed test separate from the satisfaction test. Record how long it takes from deciding to try something to having a result you can inspect, including upload and download when relevant. Fast generation can preserve a train of thought. It can also increase the number of almost-right images you feel obliged to examine. The useful result is a faster decision that leads to a usable asset. A larger folder is not itself evidence that your work has improved.

Recraft has an official ComfyUI integration, and its role is to call the external API. A node on your canvas does not imply that the image model is running on your Mac or on a graphics processor you rented. For an initial experiment, I would keep the connection simple and verify which current models the integration exposes before relying on it. A short direct API call can be enough. A specialist generator should not require you to reorganise every other part of the workshop.

My recommendation is to begin with Styles and editable vector output, because that combination tests Recraft's most distinctive potential contribution to your toolkit. Give Flash its own small trial for fast exploration. Keep your established reference-editing and local training options available for jobs where identity, control of the model, or repeatability matters more. For factual diagrams and publication data, make the visual decisions from verified inputs. Recraft can supply useful ingredients around that work without being asked to invent the facts.

Recraft is trying to make machine-generated images participate in a design process. That ambition is worth taking seriously. The evidence that would persuade me is modest and demanding at the same time: an approved set of assets, files that remain easy to work with, and fewer minutes between the brief and the finished use. You do not have to believe in the company's entire future to find one valuable job for its models today.

Pär, give Recraft a small, real assignment and a clear finish line. Let the exported artwork argue its case. If it earns the next assignment, you have learned something useful. If it does not, the experiment has still answered the question without becoming a platform migration. Claude can have an excellent episode too. The actual competition is against the time between having an idea and shipping something that works.