Seedream 5.5 Held Image Details Through 100 Edits as GPT Image Models Drifted

In a Nov. 9 comparison, ByteDance’s Seedream 5.5 kept more details of a source image intact through 100 successive edits than OpenAI’s GPT Image 2.5 and GPT Image 2.6. The sequence tested how details changed when each generated image became the next round’s input.

The test began with a photo of a wrist wearing a Casio G-Shock watch. Each prompt asked the model to advance the seconds shown on the watch by one, and the output from each round became the starting image for the next. All three models advanced the seconds through the sequence.

Seedream kept the wrist, lettering on a snack package and background details visually consistent from the first edit through the 100th. The two GPT Image models showed more changes elsewhere in the picture as the repeated edits continued.

GPT Image 2.5 began showing red marks on the arm at about the eighth edit. The package text became garbled around edit 14. By the 50th round, only the numbers on the watch face remained legible; by the 100th, the picture had turned into multicolored blocks.

GPT Image 2.6 changed in a different way. Its colors grew more saturated starting around edit 12. By about the 85th round, the original scene had been replaced by a product photo: the wrist, package and background were gone, and the watch had changed. Its date display read “SU 6-30,” while the time remained correct.

The comparison covers one photo and one repeated instruction, so it does not establish an overall ranking of the models. Some commenters questioned whether the test was fair, arguing that Seedream appeared to change only the requested area while GPT redrew more of the image. Others said that starting each GPT edit from the original photo would avoid accumulating changes across rounds.

The exercise examines Seedream 5.5’s advertised 100-edit capability by tracking what happens when each result feeds the next round. TokenPost previously covered Seedream 5.5’s 100-round image-editing capability.