First Look at Google's New Imagen 2 & Image FX Interface!

Google’s Imagen 2–powered “Image Effects” interface in the AI Test Kitchen stands out less for raw image quality and more for how it turns prompting into an interactive, exploratory workflow. Users can type a prompt, then click blue-highlighted words to swap attributes via dropdown suggestions—switching styles (photo vs drawing), composition cues, and descriptors in a way that feels more like steering an image than starting over from scratch. The result is a fast loop for iterating on “what the model thinks you mean,” with outputs that often land in strong photo-real territory.

Early generations look genuinely polished for a general-purpose image tool, with close attention to photographic details and a realism level the creator compares favorably against Midjourney and “Dolly 3” in certain cases. The interface also supports a “seed” control: locking the seed lets users change one element at a time (like “massive wave” to “small wave,” or swapping a “tabby cat” for a “purple cat”) while keeping the underlying variation stable—so changes behave more like edits than full rerolls. That makes it easier to understand cause-and-effect in prompting and to refine a concept without losing the overall framing.

Where the system draws sharp lines is policy enforcement. Prompts containing certain words—such as “battle,” “mediocre,” “animated,” or even seemingly unrelated terms—get blocked or altered, and the creator repeatedly hits cases where the model refuses or substitutes different words. The strictness is framed as a tradeoff: tighter safety controls can limit experimentation, and the interface sometimes blocks creative directions even when the intent is harmless. Still, the tool proves more permissive in some high-interest categories, especially famous characters.

A recurring “best use case” is generating recognizable characters in realistic scenes—particularly when paired with brands and settings like McDonald’s. The creator demonstrates Sonic, Bowser, and Mario-like figures eating McDonald’s items, noting that the character likeness and brand coherence are often strong, even when other prompts struggle. Attempts at more abstract or stylized directions (like “animated” or certain cinematic battle concepts) run into refusals or weaker results.

Text generation is treated as a mixed bag: the model can produce readable text in some cases, but it doesn’t consistently match the best dedicated text capabilities, and it may hallucinate unexpected phrases. The creator also points out common image-generator failure modes—like incorrect subject transformations (a man becoming a frog) and occasional anatomical oddities (hands or arm counts)—alongside moments where fixes work when the prompt is reframed more directly.

Access is described as available through the AI Test Kitchen’s Image Effects launch page, with availability varying by country (the creator reports the United States works). Overall, the interface’s biggest value is the prompting experience—especially the dropdown-driven attribute swapping and seed-based iteration—while performance remains uneven depending on prompt type, safety constraints, and the complexity of the requested scene.