Introducing Recraft V4: Design Taste Meets Image Generation

TL;DR

Recraft V4 is a commercial text-to-image model from Recraft, launched 2026-02-17, positioned as a “designer-tuned” alternative to broad-preference general models. The product comes in four variants — V4 / V4 Pro at 1024² and 2048² resolution, plus V4 Vector / Vector Pro that output editable SVGs directly from text. The blog frames V4’s contribution as “design taste”: composition, color relationships, typographic integration, and material realism rather than raw photoreal fidelity, with side-by-side comparisons against Nano Banana Pro and GPT High on portrait, typography poster, and product-shot prompts. No technical details (architecture, training data, parameter count) are disclosed; the post is a launch announcement plus qualitative comparison gallery.

Key claims

V4 is a ground-up rebuild optimized for “visual taste, prompt accuracy, and output quality that holds up at any size,” developed in collaboration with designers rather than for broad user preference [§ “Introducing Recraft V4”].
The model ships in four SKUs: V4 (1024², ~10s, lower cost), V4 Pro (2048², ~28s, higher cost), V4 Vector, and V4 Vector Pro; all share the same creative capabilities and design aesthetics, differing only in resolution/scale [§ “V4 is available in two versions”; § “Try V4 in Recraft Studio”].
Recraft claims to be “the only AI model that generates editable vector files from a prompt” — actual SVGs with structured layers and clean geometry, exportable to web/print or any vector design tool with no tracing or conversion step [§ “Recraft V4 Vector and V4 Pro Vector models”].
Qualitative claim on photorealistic portraits: V4 produces “cinematic frames with distinct emotional atmosphere” while Nano Banana Pro is “technically compliant, yet feels less expressive” — both models hit the prompt’s content requirements [§ “Go beyond photorealistic stock”].
Qualitative claim on long-prompt typography posters: V4 treats typography as a structural composition element (letters “compressed inside the cavity, pushing toward the frame edges”) rather than a decorative overlay, while Nano Banana keeps inflated letterforms centered and self-contained [§ “Generating images with text”, “OVERTHINK” comparison].
Qualitative claim on multi-product mockups: V4 follows explicit count constraints (“three matte sage-green bottles and one matching jar”) and preserves consistent branding across all units, while Nano Banana Pro gets the count wrong and misses branding on one bottle [§ “Describe what you want”, BOTANICA LAB comparison].
All variants — Standard, Pro, Vector, Vector Pro — are available to every user including the Free plan [§ “Try V4 in Recraft Studio”].

Method

The post is a product launch and discloses no technical details: no architecture (the existence of a separate “Vector” SKU strongly implies a separate vector-output model rather than a single shared backbone, but this is not stated), no training-data description, no parameter count, no objective/loss, no human-preference benchmark numbers. The methodological framing is entirely on the target rather than the mechanism: Recraft says V4 was “developed in close collaboration with designers” and “tuned around design aesthetics and professional expectations,” with the implication that the preference signal used for tuning is curated by professional designers rather than crowd-sourced. The body of the post is a qualitative comparison gallery — pairs (occasionally triples) of V4 Pro / Nano Banana Pro / GPT High outputs on identical prompts spanning macro portraits, editorial fashion, long-prompt typography posters, vector branding, and product mockups. The comparisons are cherry-picked by Recraft.

Results

No quantitative benchmarks. The headline results are qualitative comparisons on chosen prompts:

Portraits. On macro skin-detail and windy-portrait prompts, V4 Pro and Nano Banana Pro both satisfy prompt content; Recraft argues V4 has stronger “atmosphere and visual intent.”
Simple-prompt portraits with action (e.g. “Close up of a model doing her own makeup”): V4 leans editorial/campaign; Nano Banana and GPT High lean stock/commercial.
Long-prompt typography posters (the “OVERTHINK” 3D-typography emerald-green poster prompt, and the “TIDELINE” foggy-marsh editorial poster): V4 integrates typography spatially with the scene (letters press against the cavity frame; title bridges above/below waterline); Nano Banana renders typography as a centered overlay.
Multi-product brand mockup (BOTANICA LAB hair-mask line): V4 follows the “three bottles + one jar” count and applies consistent labels; Nano Banana Pro gets the count wrong.
Product still-life (papaya + green box): V4 keeps the prompt’s two subjects centered and dominant; Nano Banana crops awkwardly.

Latency and resolution numbers are given for the SKUs (V4 ≈ 10s @ 1024²; V4 Pro ≈ 28s @ 2048²). No FID, no CLIP score, no human-preference Elo, no comparison to other text-rendering benchmarks.

Why it’s interesting

Recraft’s positioning is the noteworthy bit, not the technology (none of which is disclosed). They explicitly target a niche the frontier general-purpose models neglect — production design work where typography integration, brand consistency across mockups, and editable vector output matter more than raw photoreal fidelity. The vector SKU is the genuinely differentiated artifact: a model that outputs structured SVG directly from prompt would, if it works as claimed, sidestep the standard raster→trace pipeline and be useful as a primitive for any logo / icon / illustration generation product. For Luma the relevant signal is that “design taste” as a tunable preference axis (presumably via designer-curated RLHF or DPO data) is a viable product story distinct from generic photorealism — adjacent to ERNIE-Image-Aes: Robust Image Aesthetics Scoring with Balanced Category Generalization‘s point that aesthetic scorers carry systematic category biases, and that the choice of preference data fundamentally shapes what a generative model learns to value. The all-qualitative, cherry-picked comparison format is a reminder that public model launches now routinely skip benchmarks entirely; without independent eval there is no way to validate the typography/vector claims.