Users report GPT Image 2.5 is a downgrade from 2.0 in realism and instruction-following
Cubic_Inequality504 · reddit · 2026-10-18
A detailed Reddit post argues ChatGPT's Image 2.5 is a clear regression from Image 2.0, especially for photorealistic work that casual meme-generation users don't notice.
Key complaints:
- CCTV-style outputs went from clean and realistic to a "pixel marmalade" of frayed textures
- No professional/candid photo preset in 2.5 escapes the pixel artifacts
- Worse understanding of human proportions, clothing, poses and expressions; attached-face generations produce absurd results and aspect ratios come out wrong (16:9 or 9:16) unless spelled out — the author suspects covert censorship
- With GPT-5.6, iteration devolves into a loop of "you're right, I got this wrong" followed by new errors, where a single attempt sufficed on 2.0
The author notes similar complaint threads have appeared over the past month.
More from Models
- Text-only AGI skeptics: all modalities are needed for general intelligence — Angaisb_ · 2026-10-18
- Blogger claims OpenAI quietly cut usage quotas after resetting them — lxfater · 2026-10-18
- ChatGPT Pro spotted auto-creating subagents, but thinking capped at minutes — thomasahle · 2026-10-18
- DeepSeek-V4.1-Flash-1017 GGUF appears on HF via Unsloth, no official announcement yet — MundanePercentage674 · 2026-10-18
- ChatGPT guardrails flag fully-clothed family photos, frustrating photo restoration user — johnellenbe · 2026-10-18
- Hugging Face founder: $1B of training compute shared free on the Hub, up 25x since 2023 — Thom_Wolf · 2026-10-18