🎨 Reve 2.0 is now in Chat AI: a model that does exactly what it says in the prompt
Most neural networks fill in some details for you. Reve, on the other hand, reads the request literally and assembles the frame according to your rules
- It follows the prompt precisely: several objects, specified colors, and positioning in the frame (“red cube on the left, three blue balls on the right”). Where others confuse details, Reve holds the composition
- Works flawlessly with text and typography: lettering, logos, posters, packaging, and Cyrillic. The result is a clean design without any letter confusion
- Editing with words after generation: generate the image, then edit the text (“change the background to night,” “remove the person on the right”). No need to regenerate everything
What is the strength compared to others:
GPT Image 2: A powerful all-rounder with “common sense,” but less accurate in strict compositions
Nano Banana: The king of character consistency based on reference, but weaker in text and complex scenes
Reve 2.0: Maximum prompt accuracy plus perfect text in the frame. A choice when it’s important to get exactly as intended.
For this example, we created three generations of the same prompt, but used different neural networks.
A cozy loft-style coffee shop, viewed from the inside at a slight angle from above. In the foreground on the left is a wooden table with three white ceramic cups with heart-shaped latte art and one red cup without a design. On the right, an open book with a blue cover lies on the table, with gold-framed glasses perched on it. In the center, a red-haired barista in a green apron is making coffee behind the counter. Behind him, “REVE COFFEE - 250₽” is written in white chalk on a black chalkboard. To the left, a pink neon sign hangs on the brick wall with the inscription “Open 24/7.” Two tabby cats are on the floor: a gray one is sleeping, a ginger one is walking toward the table. Rain is falling outside the window to the right, and the evening light reflects the warm yellow light of Edison bulbs in puddles on the floor. Photorealism, soft depth of field, cinematic lighting
As you can see, Reve 2.0 followed the instructions 100% And when asked to replace a single element in an image, the neural network followed the instructions precisely, as it generates the image layer by layer and can change them without any problems.
The only limitation: until the model accepts an image as a reference, it works via a text prompt and subsequent editing. This feature will be available in upcoming updates.
Try out the new product in a format convenient for you: 👇
🌐 Website
Application (on Android)
Application (on iPhone)
