Hello 
Back with an API release thread for the month of July!
The two most advanced image upscalers are here. Wonder 3.5 for upscaling almost any image, and Bloom 2 for upscaling and adding detail to AI-generated images.
New Models
Bloom 2 is our creative upscaler for native GenAI content — AI artwork, illustrations, and CG renders. It enhances image quality while introducing new detail, texture, and visual elements, with the range to either subtly refine your source or substantially reimagine it.
A Creativity control with five levels — Subtle, Low, Medium, High, and Max — determines how strongly Bloom 2 transforms the source, from restrained enhancement all the way to bold reinterpretation with newly generated detail. Keep in mind that the higher the level, the more the model generates — and the more likely it is to introduce hallucinations or artifacts. Subtle and Low stay closer to your source, while High and Max take more creative liberty. It also adds optional Face Recovery to improve facial and skin detail and cut the plastic look common in AI images, plus Color Preservation to hold your source colors while new detail is introduced.
One of the biggest improvements in Bloom 2 is a meaningful reduction in hallucinations. Where earlier creative upscaling could introduce artifacts like extra fingers or distorted anatomy, Bloom 2 handles subjects more reliably — producing cleaner, more coherent results with far fewer of those unsettling mistakes.
What’s new in Bloom 2:
- Improved creative upscaling for GenAI images with new realistic detail and texture
- Fewer hallucinations and anatomical artifacts like extra fingers or body distortion
- Enhanced skin and facial detail for more natural-looking subjects
- Creativity control with Subtle, Low, Medium, High, and Max levels
- Optional Face Recovery and Color Preservation for greater control over final results
We want to see what you create — especially on AI images where earlier models left plastic skin, flat detail, or anatomical glitches.
Wonder 3.5 builds on Wonder 3 with higher detail, greater precision, and more natural realism. Same generative upscaling model you already use, improved across the board.
You get finer detail with fewer repetitive patterns, better handling of visual noise, and stronger results on the heavily degraded and compressed images that are hardest to recover. The biggest gains show up on low-resolution and compressed photos — more transformative enhancement with clearer detail and sharper text — and on structured graphics like tables and charts, where it’s noticeably more precise. It still holds strong across portraits, landscapes, wildlife, and complex textures.
What 3.5 improves over Wonder 3:
- Higher-quality results with increased detail and improved realism
- More transformative enhancement for low-resolution and compressed photos, with better detail and text clarity
- Fewer repetitive patterns and improved handling of visual noise
We recommend referencing the linked model pages above to view full documentation that includes model descriptions, pricing, endpoints / parameters, and output examples.
As always, we’d love to hear your feedback, questions, and ideas! If you’d like to get involved, feel free to leave comments here in our community forum or in our Developer Discord!
Changelog
-
Launched two new image models
-
Bloom 2
-
Wonder 3.5
-
For any issues, reach out to support@topazlabs.com. Enterprise inquiries can be directed to enterprise@topazlabs.com.
Aiden Warnke
Product Manager
Enterprise & API
