We spoke with the CEO of Freepik

Freepik CEO details massive AI pivot and growth strategy

Tencent unveils real-time interactive video generation

New Meta prototype headsets at SIGGRAPH

Mysterious Nano Banana AI image editor sparks speculation

Cuenca panicked when DALL-E 2 launched in 2022, realizing it would make Freepik's entire stock image business obsolete within years.

The company quickly integrated Stability's open-source models as a starting point, then built custom interfaces and workflows on top.

They recently eliminated the credit system entirely for Premium Plus and Pro subscribers, offering unlimited image generation after calculating worst-case scenario costs.

Cuenca rejects viewing Freepik as just a "marketplace of models," instead positioning the UI and user experience as the real product.

The team carefully tests and filters which AI models make it onto the platform, only adding ones that excel in specific areas like anime or particular visual styles.

They're building toward a conversational interface where users describe what they want without needing to select specific models.

Cuenca sees a split emerging between end-to-end AI creation for measurable campaigns like marketing and precision tools for creative projects like films where you "have one shot to be good."

The platform's fastest-growing user segments are marketing professionals, Hollywood VFX artists, and product placement teams, replacing expensive photo shoots.

Watch the full interview for deeper insights on legal challenges, output moderation, and Freepik's expansion into project management features.

The platform automatically converts any video or audio file into structured, readable blog posts using speech recognition and natural language processing.

You can customize the AI output with specific prompts to match your brand voice, tone, and style preferences for consistent messaging.

Visual content gets handled automatically with screenshot extraction from key video moments, plus options for stock images and AI-generated graphics.

YouTube integration detects new uploads and converts them into draft or published blog posts completely hands-off.

Multi-language support lets you repurpose content for global audiences without translation overhead, while built-in SEO optimization boosts search visibility.

Yan-Sim delivers high-fidelity simulation of interactive game videos at 1080p resolution and real-time 60fps performance using a diffusion model adapted into a causal architecture.

Yan-Gen enables versatile generation of diverse interactive video content from text, visual, and action-based controls using a multimodal diffusion transformer backbone.

Yan-Edit supports multi-granularity video content editing through text-based interaction, handling both structural changes like adding objects and style changes like altering colors and textures.

The modular design lets you improve or swap each component independently, making Yan a general-purpose backbone for different world modeling tasks across game genres and edit the world as you're in it.

Tiramisu delivers 90 pixels per degree resolution (3.6x sharper than Quest 3) with 1,400 nits brightness and HDR-like contrast, but sacrifices field of view to just 33° x 33°.

Boba 3 flips the script with a massive 180° horizontal field of view that covers 90% of human vision, compared to Quest 3's 46% coverage.

Both prototypes use existing consumer components where possible but remain bulky research devices that require high-end PCs to run.

The demos run real-time Unreal Engine 5 graphics with NVIDIA DLSS to handle the massive performance demands of either ultra-high resolution or ultra-wide rendering.

Meta calls this progress toward the "visual Turing test" where you can't tell VR from reality, though neither prototype solves all the technical challenges simultaneously.

The model excels at natural language editing where you can modify photos by typing instructions like "change the person's facial expression" or "adjust the background lighting."

Early testers report getting clean, professional results on their first try, with the model maintaining character consistency and preserving the original image's overall scene context.

AI researchers speculate it could be Google's latest image model based on its sophisticated performance, though Google hasn't confirmed any connection to nano banana.

You can test the model yourself through LMarena, where it's being compared head-to-head against other AI image tools.