Imagen 4 is Revolutionizing Image Generation Again

Simply put, Imagen 4 is both versatile and high-quality enough that it blows all competition out of the water. While it lacks the editing capabilities of Flux Kontext or GPT-4o, the raw capability of the model is totally unmatched. From a wide range of styles to extreme graphical fidelity to the highest level of prompt adherence we have seen, the model impresses at every step.

Below is a showcase of images we made with Whisk, Google’s tool for generating and animating images.

As we can see, Imagen 4 is incredibly versatile. From everything from photorealism to more obscure art styles like MS Paint, Imagen 4 seems to know everything required to generate an accurate representation of your textual input. It excels in both prompt adherence and writing, with the latter being completely unmatched by any competitive models. The only limitation is the user’s creativity and the quality/depth of their prompt.

So how do we prompt Imagen 4 in a way that gives us the best results? Old tricks like adding the art style to the end of the prompt and using enhancing language still work, and even better than before thanks to the awesome prompt adherence on the model.

But Whisk lacks something we have found essential when using other commercial image generation tools: a tool to enhance our prompts. This is where we recommend using a commercial or open-source LLM. Plug your prompt into an advanced model and ask it to expand the scope of your prompt for an image generation task. Then, directly edit the now expanded prompt to fit the vision you have. Such editing can help elevate your image like the example shown above.

In our experiments, we found Imagen 4 to outperform the competition, qualitatively, in nearly every head to head comparison we made. Specifically, we found that the washed out color of the GPT-4o images made them inferior to the sharp colors of the other models, that Reve Halfmoon lacked the prompt adherence of Imagen or GPT-4o, and that HiDream couldn’t handle the writing task nearly as well.

That being said, it wasn’t always the case, as we can see in the bottom row example. Each model has strong capabilities in each of the tasks we tested them on. In the interest of transparency, none of these choices were specifically chosen for their fidelity and prompt adherence, we just used the first result from each generation. For example, the Imagen 4 result for the last test had the best prompt adherence but poor writing quality compared to GPT-4o and Reve Halfmoon. Nonetheless, we found Imagen-4 was still consistent enough to always warrant using it over competitive models in any situation the license approves of.

Imagen 4 is a really awesome and powerful image generator model. Not only is it far ahead of competition in prompt adherence, color quality, and graphical fidelity, but it is on par with complicated capabilities like writing that require massive VLM’s to really mimic. We are very impressed with the results of Google’s research here, and look forward to testing other efforts by Google DeepMind like Veo 3 and Gemini Diffusion, as well.