Image

Recraft V4: how to write prompts the model actually understands

Recraft · Updated:

Recraft V4 is the only AI model that produces true editable SVG with structured layers. Accurate text rendering, real design taste, two prompt modes (interpretive 3–6 words and architectural with full structure), 100+ preset styles via parameter, plus negative prompts and I2I support.

What Recraft V4 does well

The flagship feature is true editable SVG with structured layers, not a rasterized fake. It's the only AI model on the market that delivers this, which makes it indispensable for logos, icons, packaging, and any branding work that needs a vector source.

Two tiers: V4 Standard (1024×1024, ~10s/image, 60 credits) and V4 Pro (2048×2048, ~28–30s/image, 175 credits). SVG generation runs 15s on Standard and 45s on Pro with noticeably higher quality. Recraft is #1 on Hugging Face Text-to-Image Arena, with a Figma plugin and MCP integration. Output formats: SVG, PNG, JPG, PDF, TIFF, WebP, Lottie.

  • True editable SVG with layers (unique feature)
  • Accurate text with proper kerning
  • 100+ preset styles via the `style`/`style_id` parameter
  • Inpainting, outpainting, vectorization, background removal, upscaling
  • Exploration mode — 8 variations from one prompt

Two prompting modes

Recraft V4 supports two fundamentally different approaches. Interpretive (3–6 words) — the model makes its own decisions about composition, color, and tone: «Retro space station», «Minimalist coffee logo». Good for idea exploration and exploration mode.

Architectural (long, structured) — full control. Structure from global to local: A [image style] of [main content]. [Detailed description]. [Background]. [Style description]. Order: Core concept → Background → Subject framing → Physical attributes → Secondary subjects → Lighting → Camera/depth → Mood.

Elements mentioned earlier in the prompt get higher priority. That matters: the main subject always goes in the first line.

Text rendering

Recraft V4 is one of the top models for text accuracy. One rule: exact text always in quotes inside the prompt. «A storefront sign reading "OPEN 24/7" in neon letters» yields exactly that text with proper kerning.

For branding, packaging, signage, and posters this is critical — every word should be in quotes. For bilingual text or non-standard fonts, specify the font explicitly («serif headline», «display font», «handwritten script»). Size and placement also via text: «small caption bottom-right», «bold headline centered top».

Style controls and base classes

Recraft uses a `style` or `style_id` parameter instead of describing style in the text. Four base classes: `realistic_image`, `digital_illustration`, `vector_illustration`, `icon`. Plus 100+ preset styles inside each class, plus custom styles via uploading a reference image.

Writing «in flat vector style» in the text works noticeably worse than picking `vector_illustration` via the parameter. Style is a separate entity from content; keep them separate. For branding you typically want `vector_illustration` or `icon`; for marketing banners — `realistic_image` or `digital_illustration`.

Colors are set via an array of preferred RGB colors + optional background color. For brand projects this gives exact hex control unavailable in most other models.

Common mistakes

  1. 1. Styles in the prompt text instead of the parameter

    «In flat vector style» or «as a digital illustration» in the prompt text works noticeably worse than picking the matching `style`/`style_id` parameter (`vector_illustration`, `digital_illustration`, `realistic_image`, `icon`). Style is a separate entity from content. Use the parameter for style, text for scene description.

  2. 2. Double negatives in negative prompts

    In negative prompts write what to EXCLUDE without «no». «blurry, text, watermark» works; «no apples» in a negative prompt is a double negative — the model misreads it. If you want to remove apples, just write «apples» in the negative field without a negation particle.

  3. 3. Plural ambiguity

    «Cats sitting on a windowsill» is ambiguous — the model can produce one, two, or five. Specify the number explicitly: «three cats sitting on a windowsill». This matters especially for design tasks where composition must be predictable. For icons and logos a concrete element count is critical.

  4. 4. Texture language for vectors

    For vector_illustration or icon, describing «brushstrokes», «paper texture», «soft fabric» is useless — vector graphics are built from geometric primitives, not textures. Use geometric and structural language: «clean lines», «structured layers», «balanced composition», «negative space», «modular grid». For textures, switch to digital_illustration or realistic_image.

  5. 5. Dramatic evaluative adjectives

    «Stunning masterpiece», «breathtaking», «award-winning» — for Recraft this is noise. The model values precision over hyperbole. Swap evaluative words for concrete descriptive elements: instead of «stunning logo» — «logo with central monogram, serif type, deep navy palette». Precision always beats pathos.

Before / after examples

Example 1

Before

a beautiful minimalist coffee shop logo

After

A minimalist logo for a specialty coffee shop. Central icon: a single coffee bean integrated with a wavy steam line above. Below the icon, brand name (EXACT): "BREW & CO" in geometric sans-serif. Tagline below in smaller letters (EXACT): "Slow roasted since 2018". Earth tone palette: warm brown #6B4423, cream #F5E8D3, deep black for text. style: vector_illustration. Clean geometry, structured layers, balanced negative space.

Architectural mode with full structure: icon → brand name → tagline, hex colors specified, style set via the vector_illustration parameter. Exact text in quotes with EXACT marker. Yields a production-ready SVG logo in one iteration.

Example 2

Before

meditation app icon

After

Meditation app icon, flat vector style. Central element: a stylized lotus flower with 5 petals, symmetric, soft purple to deep blue gradient on petals (#A78BFA → #4338CA). White background with subtle radial glow. style: icon. App store ready, rounded square format, balanced negative space, no text.

Concrete hex colors for the gradient, icon style via parameter, composition balance described explicitly. «No text» in the main prompt is not a negative — it's a positive instruction to leave the background empty.

Example 3

Before

three cats on a windowsill in watercolor style

After

A digital illustration of three cats sitting on a wooden windowsill. From left to right: a black cat with green eyes, a ginger tabby looking sideways, a white cat with blue eyes. Window behind them shows soft afternoon light, blurred garden in background. style: digital_illustration. Watercolor texture, soft pastel palette, painterly brush strokes. Detailed but not photorealistic.

Number of cats specified (three, not «cats» — plural ambiguity). Each cat described for predictability. Style set via parameter + watercolor texture in text for technique specifics.

Frequently asked

How does V4 Standard differ from V4 Pro?
Standard — 1024×1024, ~10 seconds per image, 60 credits, SVG generation in 15 seconds. Pro — 2048×2048, 28–30 seconds, 175 credits, SVG in 45 seconds with noticeably higher quality. Pro delivers better detail and SVG structure, which matters for production logos and complex illustration. Standard is enough for prototyping, icons, and quick idea exploration.
Which output formats are supported?
SVG, PNG, JPG, PDF, TIFF, WebP, Lottie. SVG is the headline feature: true editable vector with layers, not a rasterized fake. Lottie is for animated JSON icons (relevant for web and mobile). PDF for print materials. Most formats are available without extra credits after the initial generation.
When should I use interpretive versus architectural mode?
Interpretive (3–6 words) — for idea exploration and exploration mode, when you want the model to make design decisions. «Retro space station», «Minimalist coffee logo» — the model picks colors, composition, tone. Architectural — when you need full control: exact text, specific colors, defined composition, fixed style. Common workflow: start interpretive to find direction, then switch to architectural for the final render.
How do I work with brand colors?
Recraft supports an array of preferred RGB colors + optional background color as a separate parameter. This gives exact hex control over the palette — critical for branding. Specify hex colors in the prompt text too (e.g., «warm brown #6B4423») for extra anchoring. Custom styles via reference image upload let you lock a brand's visual language across a series of generations.
Can I get an editable SVG?
Yes, this is the headline V4 feature — true SVG with structured layers that opens in Figma, Illustrator, Inkscape for further editing. Every element is a separate path; you can change colors, move, delete. This is unique on the market: most AI models either can't do SVG or hand you «traced» SVG-as-raster with no structure.
What's not supported in V4?
Style creation (building a new style from scratch), prompt-based editing (text-prompt edits on an existing image — that's GPT Image 2 territory), image sets (generating a series with unified composition), artistic level control. For these tasks either wait for V5 or chain other models in the pipeline. Inpainting and outpainting are in V4 but are not the same as prompt-based editing.
Does Opten support Recraft V4?
Yes, the Opten extension recognizes Recraft inside recraft.ai and scores prompts against the model-specific structure: it checks for exact text in quotes, style set via parameter instead of in text, absence of double negatives in negative prompts, explicit object count, and texture language matching the chosen class. One click yields a rewrite in the right structure.

Related models

Ready to write Recraft V4 / V4 Pro prompts in one click?

  • Auto-detects the model inside its native interface
  • Scores every line of your prompt
  • One-click rewrite into the correct structure
ChromeYandex BrowserChrome / Yandex BrowserInstall extension

Pro — $2.99/month or ₽199/month · cancel anytime

Stop Guessing. Generate
On The First Try.

Install Opten in 30 seconds and score your next prompt.

Opten is a Chrome extension that scores AI prompts for the specific model. Supports 60+ image and video models — Midjourney, GPT Image 2, Kling, Sora, Nano Banana, Flux — and rewrites them in one click inside the Syntx, Higgsfield, and Freepik interfaces. From $2.99/month.

© 2026 Opten · IE Nikolai Shupletsov · Tax ID 306389672