Gemini image generation stops refining and regenerates whole scenes on follow-up edits
asm99 · reddit · 2026-09-23
A Gemini Pro user reports that iterative image editing broke a few days ago: initial generations still follow prompts, but any follow-up request to tweak lighting, poses, angles, or composition regenerates the entire scene from scratch instead of refining the existing image.
The issue reproduces on both web and mobile, persists after logging out/in and clearing cookies, and appeared without any user-side settings change — suggesting a silent server-side change. Other users running iterative image workflows may want to check whether they're affected.
More from Multimodal
- Pooling-aware training cuts retrieval vectors 4x while boosting nDCG to 97.8% of unpooled — lateinteraction · 2026-11-11
- Generating accurate 3D building models from 150 planning PDFs with new Gemini model — fofrAI · 2026-11-11
- Stanford's DyCraft Uses Agentic Decomposition to Cut Physical Errors in Long-Horizon 4D Scene Generation — StanfordUniversity · 2026-11-11
- Midjourney SREF pick: --2523103199 for loose pencil-sketch portraits — tisch_eins · 2026-11-11
- H3 Image Tips: 16MP via ComfyUI Workflow Plus Detailed Prompts Yields Best Results — gatortux · 2026-11-11
- Gemini-generated endless loop of quantum cryostats self-assembling — hapticdata · 2026-11-11