Alibaba Open-Sources Qwen-Image-2.1: One 7B Model for Image Generation and Editing

On Sept 20, Alibaba's Qwen team released and open-sourced Qwen-Image-2.1, a unified text-to-image and image-editing model, with weights on Hugging Face (Qwen/Qwen-Image-2.1) and a live Spaces demo; Comfy-Org shipped ComfyUI-ready files the same day, and Diffusers and SGLang provided Day-0 support.

Confirmed

• The vision component is only 7B parameters, a 32-layer single-stream DiT with hybrid-granularity attention; the team claims much faster multi-image inference and performance surpassing most closed-source models, positioning it as the most balanced, cost-effective model in the series
• Native 2K generation and RGBA transparency support, enabling direct transparent-background image generation
• Instruction-based editing with up to 10 reference images, pixel-level 2K editing, and identity consistency for people and products
• Local editing demo: circling three regions and using one instruction to remove a metal watch, change hair to black, and swap clothing in a single pass
• Character consistency: given a three-view character reference, the model can generate full storyboard panels; it also suits panoramas, infographics, and virtual try-on
• Per Lexius2129, texture quality clearly improves over Qwen-Image-Edit, and the model iterates on the Qwen-Image 2.0 architecture released in February, with emphasis on text rendering aesthetics and portrait quality
• Per RisingSayak, the Diffusers Day-0 support uses a novel KV-caching approach with a decoupled prefill+decode architecture
• Low deployment barrier: Lexius2129 reports 30 seconds per image on an RTX 5080 laptop; community reports note SGLang-Diffusers runs on a single RTX 4090 with about 22.7GB VRAM
• Released under the qwen-research license (research use only), supported in vLLM-Omni, ComfyUI, and Hugging Face Spaces; @akhaliq created with it on day one and the official account amplified the post
• Per Lexius2129, Qwen-Image-Bench score of 60.28 exceeds NanoBanana2.0 (59.82) and GPT Image

Why it matters

• The 7B size dramatically lowers local deployment barriers—consumer GPUs and laptops suffice—while open weights and day-0 ecosystem support (ComfyUI, Diffusers, SGLang, vLLM-Omni) make it friendly for secondary development and real applications
• Unified generation/editing, one-shot multi-region editing, transparent image generation, and three-view-to-storyboard cover high-frequency workflows in character design, e-commerce, and content production; the claimed benchmark lead over major models, if validated, would further squeeze closed-source image models

2026-09-20 ~ 2026-09-21 · 36 related posts

Full story(10 episodes)→

• Episode 1: Qwen Image 2.1 open-source model appears set for release(2026-09-18, 2 posts)
• Episode 2: Early Qwen Image 2.1 hands-on reports split: fast and good at editing, weak photorealism(2026-09-19, 7 posts)
• Episode 3: Alibaba's Qwen-Image-2.1 Announced as Fully Open-Source with Day-One ComfyUI Support(2026-09-20, 7 posts)
• Episode 4: Qwen Releases Prompt-Rewriting Models for Image Generation(2026-09-20, 2 posts)
• Episode 5: Qwen Image Model Beats Flux in Quality and Speed, Users Report(2026-09-20, 4 posts)
• Episode 6: Alibaba Open-Sources Qwen-Image-2.1: One 7B Model for Image Generation and Editing(2026-09-20, 36 posts)
• Episode 7: Qwen Image 2.1 Tested: Strong Text Rendering, Yellowish Tones(2026-09-21, 2 posts)
• Episode 8: Qwen-Image 2.1 Lands on SGLang with 2.55x Speedup from KV Cache(2026-09-21, 2 posts)
• Episode 9: Qwen-Image-2.1 GGUF Quantized Version Tops Hugging Face Trending(2026-09-21, 4 posts)
• Episode 10: Unsloth's quantized Qwen-Image-2.1 runs locally on just 6GB VRAM(2026-09-21, 2 posts)

• Alibaba's Qwen releases Qwen-Image-2.1 image generation model, weights on Hugging Face — StableLlama · 2026-09-20
• Qwen-Image-2.1 weights now available on Hugging Face — SamuelTallet · 2026-09-20
• Alibaba open-sources Qwen-Image-2.1: unified 7B gen+edit model with transparency and 10-image input — 千问大模型 · 2026-09-20
• [source] Alibaba open-sources Qwen-Image-2.1, a 7B unified image generation and editing model — Alibaba_Qwen · 2026-09-20
• Qwen-Image-2.1 demo shows three-region local edits via circle annotations in one pass — Alibaba_Qwen · 2026-09-20
• Qwen-Image-2.1 turns a three-view character reference into a full storyboard — Alibaba_Qwen · 2026-09-20
• Qwen-Image-2.1 arrives with better text rendering, portraits, and multi-view editing — Alibaba_Qwen · 2026-09-20
• Qwen-Image-2.1 released open-weights: 7B model with native RGBA layers and 10-image editing — ResearchCrafty1804 · 2026-09-20
• Qwen-Image-2.1 is out: open-weights 7B image generation and editing model — Front-Relief473 · 2026-09-20
• Qwen-Image-2.1 official blog: native transparency, up to 10 reference images, flexible local editing — LatentSpacer · 2026-09-20
• Qwen-Image-2.1 weights released for ComfyUI on Hugging Face by Comfy-Org — Peasant_Farmer_101 · 2026-09-20
• [source] Qwen-Image-2.1 lands in ComfyUI: one 7B model for 2K generation and editing — Alibaba_Qwen · 2026-09-20
• Qwen-Image-2.1 Hits Hugging Face Trending With Text-to-Image and Editing — Qwen · 2026-09-20
• Qwen-Image-2.1 is now available on Hugging Face — _akhaliq · 2026-09-20
• New model ships Day-0 in Diffusers with novel KV-caching and disentangled prefill+decode — RisingSayak · 2026-09-20
• Qwen Image 2.1 launches: 7B image generation and editing model with up to 10 reference images — multimodalart · 2026-09-21
• Qwen-Image 2.1 lands Day-0 support in ComfyUI — Lexius2129 · 2026-09-21
• Qwen Image 2.1 surfaces in early look at Alibaba's image model update — petrusenko_max · 2026-09-21
• Qwen-Image 2.1 open weights land with day-0 SGLang support, runs on one RTX 4090 — ying11231 · 2026-09-21
• Qwen-Image-2.1 lands on Hugging Face with open weights on day one — Alibaba_Qwen · 2026-09-21
• Qwen-Image-2.1 demo hits Spaces: 7B single checkpoint for generation and editing with 10 reference images — Alibaba_Qwen · 2026-09-21
• Qwen-Image-2.1 open weights: 7B image model runs on RTX 3090, 2K output in ~21s — 量子位 · 2026-09-21
• Qwen-Image-2.1 model card: mixed-granularity attention, prefix KV reuse, quickstart code — AdinaYakup · 2026-09-21

13 near-duplicate retellings: multimodalart · Xianbao_QIAN · 机器之心 · linoy_tsaban · RisingSayak · airesearch12 · AIandDesign · Lexius2129 · gaganghotra_ · 智东西 · AdinaYakup · Arindam_1729 · RisingSayak