Multi-turn editing
Unedited regions preserved at pixel level, shown across 100 edits
Single and multi-reference image editing
Transparent output
From a prompt or reference images
With exactly one reference image
Text rendering
0.4% severe and 1.7% minor text corruption in ByteDance's test
Rates not given; ByteDance says 5.5 is substantially lower than the previous version
Layer decomposition
Base image plus up to 99 layers, optional text recognition
Base image plus up to 16 layers
Size and references
1K, 1.5K, 2K or exact pixels; up to 10 reference images
1K, 1.5K, 2K or exact pixels; up to 10 reference images
Pricing
Per image by size, plus each reference image
Per image by size, plus references after the first