Does DALL-E 3 Have Competition? Open Source GPT-4 Vision & more! | AI NEWS

Adobe is rolling out a major upgrade to its Firefly image generator, positioning the new Firefly Image 2 model as a serious alternative for creators who care about realism, controllability, and production-ready output—especially inside Adobe’s own workflow. Firefly’s outputs are described as commercially usable, and prompts are claimed to work across 100+ languages. The update lands in the Firefly web app and also ties into tools like Photoshop (Generative Fill and Generative Expand), Illustrator (generative recoloring), and a standalone Firefly web experience meant to compete with mainstream text-to-image systems.

The headline change is Firefly Image 2, which Adobe says improves creator control and image quality, along with faster generation and easier use. Users can toggle between “image 2” and the older “image 1” model. Beyond the model swap, the web interface adds new controls aimed at making outputs look more like real photography: a “visual intensity” slider to shift toward more artistic or more photographic results; a style strength slider; and a photo-settings panel that exposes camera-like parameters such as aperture (down to f1.2 and up to f22), shutter speed (from 1/12000 to 10 seconds), field of view (roughly 14mm to 300mm+), plus negative prompting. The interface also supports “style” transfer by uploading a reference image and applying its look to a new generation.

Community testing and side-by-side comparisons suggest Firefly’s strengths cluster around fine-grained realism and detail—particularly in hair, skin texture, and lighting—while other models may win on raw prompt adherence or specific failure modes. In one set of comparisons, Dolly 3 is described as more literal and stronger at text generation (Firefly is portrayed as weak or absent for text), while Midjourney is credited with strong results but sometimes shows background or object oddities. Firefly is repeatedly praised for looking “photo-like,” including wrinkles and hair strands, though it’s also shown producing occasional malformed artifacts and “bug” images where coherence breaks.

A standout feature in the hands-on demos is style transfer: applying the style of an uploaded image (including the creator’s own channel logo) to the same underlying concept. The results range from convincing hybrids (a cat-dog blend) to intentionally disturbing outcomes when a “decrepit” face image is used—framed as “nightmare fuel.” Adobe also adds guidance from team members: using built-in lighting and intensity controls can steer results toward more authentic photo aesthetics, and text alignment improvements can prevent common issues like extra fingers.

Pricing is presented as credit-based: a free tier includes $25 monthly generative credits, while a premium tier offers 100 monthly credits for $5/month—though the channel host suggests heavy Adobe users may effectively get unlimited generations through existing subscriptions. The broader competitive landscape remains unsettled: Firefly is not portrayed as a Dolly 3 killer overall, but it’s framed as a strong option for product photography and realism-focused workflows.

The update comes alongside another major theme in AI image generation: safety restrictions. Dolly 3’s image generation inside Microsoft’s Image Creator is described as tightening, with inconsistent enforcement—sometimes blocking prompts that should be allowed, while other prompts slip through. The segment also touches open-source vision alternatives (LLaVA on Hugging Face), claims about GPT-4 Vision decoding redacted NASA text with near-100% accuracy in one test, and a broader AI trajectory discussion that includes talk of GPT-4.5 and timelines for artificial superintelligence.