What's new

We are expanding our visual content capabilities and adding support for the Google Nano Banana Pro model.

New features are now available to you directly in the chat interface:

• Image generation: create graphics from scratch using text descriptions.
• In-dialog editing: make changes and refine images directly while chatting with the model.

This update allows you to work with images faster and more conveniently without leaving the familiar chat window.

• Improved color correction: colors are more accurate, cleaner, and more natural.
• Increased detail: finer elements and textures are rendered more thoroughly.
• Updated post-processing algorithms: images look higher quality right out of generation.
• Reduced generation cost.

• Faster generation
• Fewer artifacts
• Reduced cost
• Improved tag handling and updated documentation

• Next-level photorealism: significant reduction of the "plastic" look and artificial skin smoothness that characterized earlier versions.
• Ultra-detailed textures: high-precision rendering of complex natural elements — from skin pores and wrinkles to animal fur and water ripples.
• Advanced text rendering: the model correctly generates long captions, complex layouts, and multi-line text in multiple languages.
• Deep prompt understanding: improved interpretation of complex instructions, including the spatial arrangement of objects and stylistic nuances.

• Improved prompt fidelity and greater style versatility.
• Test users report increased detail and realism.
• Significantly improved anatomy and compositional consistency.
• Improved consistency: many issues with facial "drift" and background distortion during partial image edits have been fixed.

• Improved image quality
• Seams when stitching together multiple 5-second videos are less noticeable
• Improved human and animal anatomy
• Great for anime

🌟 What's new?

• Up-to-date dataset: The model's knowledge base has been updated, data cutoff — last month.
• Improved lighting: Faces are noticeably brighter; fixed an issue with overly deep shadows.
• Composition stability: Fixed a bug where the model often produced "collages" (multiple angles in a single image).
• Prompt adherence: Slightly improved accuracy in understanding text prompts.
• Character refinement: Characters introduced in the previous version are now trained more deeply and produce higher-quality results.

⚠️ Known issues

Although overall stability has increased, the update introduced a few caveats:

1. Prompt sensitivity: The model has become more demanding of prompt quality. Short prompts or poor structure now more often lead to unsuccessful results.
2. Eye color: Rare cases of the eye color not matching what was specified in the prompt may occur.
3. Style inconsistency: Due to weight merging to support third-party LoRAs, the style can vary, especially when generating specific characters.
4. Character limitations: Some new characters (e.g., Ye Shunguang from ZZZ) still require many tags and currently work best in portraits (close-up headshots).

🛠 Technical usage recommendations

For best results, it's recommended to follow these settings:

• Prompt composition: For a single character, use the tag combo 1girl, solo.
• SFW filtering: To avoid 18+ content, add the word nude to the negative prompt (Negative Prompt).

• Adjusted the model's default overall style
• Improved visual clarity and accuracy of character rendering

• Improved speed
• Reduced usage cost
• Improved image generation quality (realism, detail)

• Support for audio. Thanks to a unified process, lip-sync and the rhythm of actions in the frame automatically adjust to the sound. If a character speaks or sings, the footage will match.
• Greatly improved preservation of characters' and objects' appearance when changing frames or angles. This resolves the problem of faces "flickering" in long scenes.
• Understanding of complex prompts improved by ~15%; the model follows the script more accurately.
• Movements are smoother and more physically accurate thanks to better understanding of depth and perspective.