Qwen-Image-2512 — December upgrade with realistic humans, finer textures, stronger text rendering (Alibaba Qwen)

TL;DR

Alibaba Qwen announces Qwen-Image-2512, a December upgrade to the Qwen-Image foundation model. The release reduces the “AI look” in human generations with richer facial detail, sharpens natural textures (landscapes, water, fur, materials), and improves text rendering accuracy and layout in text–image composition. In 10,000+ blind rounds on AI Arena, the model is positioned as the strongest open-source image model while staying competitive with closed-source systems.

Key claims

Qwen-Image-2512 dramatically reduces the “AI look” on humans, with richer facial details vs the prior Qwen-Image release [tweet body].
Natural textures (landscapes, water, fur, materials) are sharper than the previous version [tweet body].
Text rendering has better layout and higher accuracy in text–image composition [tweet body].
In 10,000+ blind AI Arena rounds, Qwen-Image-2512 ranks as the strongest open-source image model and stays competitive with closed-source systems [tweet body].

Method

The tweet is a release announcement, not a technical report; no architecture, training-data, or evaluation details are disclosed. It positions Qwen-Image-2512 as a December incremental upgrade to the 20B-parameter Qwen-Image base (see Qwen-Image Technical Report for the underlying model) — emphasis is on perceptual quality on humans/textures and text-rendering accuracy, both areas the original Qwen-Image already highlighted as differentiators.

Results

Win-rate position: strongest open-source image model in 10,000+ blind AI Arena rounds; “competitive with closed-source systems” — no absolute Elo or pairwise numbers given [tweet body].

Why it’s interesting

Qwen-Image-2512 is the next checkpoint after the November Qwen-Image-Edit-2511 release teased by Qwen-Image-Edit 2511 teased — bdsqlsz leak ('next week') and distilled into 4-step LoRAs by Qwen-Image-Edit-2511-Lightning — Step-Distilled 4-Step LoRA for Qwen-Image-Edit-2511, extending Alibaba’s monthly cadence on the Qwen-Image line (Qwen-Image Technical Report, Qwen-Image — Image foundation model with text rendering and editing (QwenLM/Qwen-Image), Qwen-Image-Edit-2509 — multi-image editing and enhanced consistency (Qwen)). The headline framing — better humans, better textures, better text — tracks the same three axes Qwen-Image emphasized at launch, suggesting Qwen is iterating on the same model spec rather than pivoting architectures.