Hi,
I am seeing what appears to be a regression in Google AI Studio when using:
Model: gemini-3.1-flash-image
Output format: Images & text
Thinking level: High
Product: Google AI Studio Prompt UI
Until around October 2, 2026, multi-image prompts could produce an interleaved Thoughts trace containing both:
- textual reasoning / planning
- intermediate generated images
The intermediate images were displayed directly inside the expandable Thoughts block, interleaved with the model’s text reasoning.
Since approximately October 2, this behavior has changed.
Current behavior:
- the model still produces a long textual reasoning sequence describing multiple image-generation steps;
- the reasoning often explicitly refers to generating, checking, or continuing with successive scenes/images;
- however, the corresponding intermediate thought images are no longer displayed in the Thoughts block, or only a single image appears;
- forcing images to be emitted explicitly as final model output images can sometimes make them appear in the final response, which suggests that image generation itself may still be occurring internally.
This looks like a regression in how intermediate image parts are exposed or rendered in the AI Studio UI, rather than simply a prompt-following issue.
Expected behavior:
Thought images returned/generated during image reasoning should remain visible inline in the Thoughts panel, as they were previously.
Actual behavior:
Thoughts now mostly contain text only, even during prompts where the model’s reasoning clearly proceeds through multiple image-generation stages.
Impact:
This breaks existing multi-image workflows that relied on the visible intermediate image parts for inspection, continuity checks, and retrieval.
BEFORE:
text thought
→ intermediate image
→ text thought
→ intermediate image
→ …
AFTER:
text thought
→ text thought
→ text thought
→ no corresponding intermediate images
Additional observation:
The model’s textual reasoning frequently behaves as if multiple images are still being generated sequentially, even when those images are no longer visible in the Thoughts panel.
Because of that, I suspect the regression may be in presentation / response-part handling rather than in image generation itself, but I cannot confirm the internal implementation.
Could you confirm whether:
- intermediate thought-image rendering was intentionally changed in AI Studio, or
- this is a regression in the UI / response handling?
Thanks.