Qwen-Image — Image foundation model with text rendering and editing (QwenLM/Qwen-Image)

TL;DR

QwenLM/Qwen-Image is the umbrella repository for Alibaba’s 20B MMDiT image foundation model and its monthly iterations. The repo tracks the original Qwen-Image (T2I, Aug 2025), Qwen-Image-Edit (Aug 2025), Qwen-Image-Edit-2509 (Sept), Qwen-Image-Layered (Dec 19), Qwen-Image-Edit-2511 (Dec 23), and Qwen-Image-2512 (Dec 31), plus the standalone Qwen-Image-2.0 announcement (Feb 2026). The Dec 23 update is the focal point of Paul’s post: Qwen-Image-Edit-2511 ships with Day-0 LightX2V distillation, vLLM-Omni inference, and SGLang-Diffusion support. Headline edit-2511 improvements over 2509 are stronger character identity, robust multi-person consistency for group-photo fusion, and improved text-in-image editing.

Key claims

Qwen-Image is a 20B MMDiT foundation model targeting complex text rendering (especially Chinese) and precise image editing [README §Introduction].
Qwen-Image-Edit-2511 (Dec 23 2025) improves identity preservation versus Qwen-Image-Edit-2509, with explicit focus on multi-person group-photo consistency [README §Showcase of Qwen-Image-Edit-2511].
LightX2V provides Day-0 acceleration for Edit-2511: 25× DiT NFE reduction and 42.55× end-to-end speedup via diffusion distillation plus inference-stack optimization, with NVIDIA, Hygon, Metax, Ascend, and Cambricon support [README News 2025.12.23].
Qwen-Image-2512 (Dec 31 2025) ranks #1 open-source on AI Arena across 10,000+ blind rounds, with substantially reduced “AI look” on human faces and finer natural textures (hair, water, fur) [README News 2025.12.31, §Showcase of Qwen-Image-2512].
Qwen-Image-2.0 (Feb 10 2026) is announced as a unified generation+editing model in a single mode with native 2K and 1K-token prompt support for infographic-style outputs; it is a separate model from the 2512 line [README News 2026.02.10].
vLLM-Omni and SGLang-Diffusion both ship Day-0 inference paths for Qwen-Image-Edit-2511 and Qwen-Image-Layered, with long-sequence parallelism and cache acceleration [README News 2025.12.23].

Method

The repo is primarily a release tracker plus reference inference code; there is no training pipeline. Each monthly variant ships as a diffusers pipeline (QwenImagePipeline for T2I, QwenImageEditPlusPipeline for multi-image edit), loaded from Qwen/Qwen-Image-Edit-2511 and analogous Hugging Face IDs. The reference inference for Edit-2511 uses 40 steps, true_cfg_scale=4.0, guidance_scale=1.0, and supports multi-image input lists (one of the 2509→2511 carryover improvements). Prompt enhancement tools (Qwen-Plus for T2I, Qwen-VL-Max for editing) are recommended for stability. Multi-GPU API server deployment ships via a Gradio harness with environment-variable-configured GPU count, queue size, and timeout. The Showcase section is the substantive evidence: side-by-side comparisons of facial detail, fur, wave physics, and slide-typography rendering between the Aug Qwen-Image baseline and Qwen-Image-2512.

Results

Qwen-Image-2512 ranks as the strongest open-source image model on AI Arena in 10,000+ blind rounds, with competitive performance against closed-source systems — exact win-rate or Elo not stated [README §Showcase of Qwen-Image-2512].
T2I-CoreBench (Nov 11 2025) reports Qwen-Image at SOTA among open-source models on real-world composition and reasoning tasks, with comparable results to closed-source systems — no numeric table on the README [README News 2025.11.07].
LeMiCa (China Unicom) achieves nearly 3× lossless acceleration on Qwen-Image inference via cache and global denoising path optimization [README News 2025.11.07].
Qwen-Image-Edit-2511 + LightX2V: 25× DiT NFE reduction and 42.55× end-to-end speedup vs. the unoptimized 40-step Edit-2511 pipeline [README News 2025.12.23].

Why it’s interesting

This umbrella repo is the canonical reference for tracking the Qwen-Image release cadence — roughly one substantive update per month from Aug 2025 through Feb 2026 — which is a useful contrast point against the slower release rhythm of Introducing FLUX.1 Kontext and the BFL Playground and Seedream 4.0 — Unified Image Generation and Editing Model (ByteDance Seed). The 2511 → 2512 → 2.0 sequence is interesting because each adds a different axis: 2511 sharpens edit-side identity/multi-subject consistency (HY-SOAR: Self-Correction for Optimal Alignment and Refinement in Diffusion Models addresses similar consistency goals via self-correction RL), 2512 closes the photorealism gap on human faces and natural textures (the same “AI look” diagnosis that motivates Bria AI Image Editing — Precision AI Image Editing Built for Production Scale), and 2.0 collapses T2I and editing into a single unified mode — joining the broader unified-multimodal trend that Learning an Image Editing Model without Image Editing Pairs, Qwen-Image-2.0, and Seedream 4.0 — Unified Image Generation and Editing Model (ByteDance Seed) all push at. The Day-0 LightX2V/vLLM-Omni/SGLang-Diffusion integration story is also notable as a maturity signal — it’s the same release pattern SGLang Diffusion — bringing SGLang's high-performance serving to diffusion models (LMSYS announcement) (SGLang Diffusion) advertised, now playing out in production.