FULL STORY

Qwen-Image-2.1: From Teaser to Open-Source Release

Alibaba's Qwen-Image-2.1 went from a leaked Coming Soon page to official open-source release on Sept 20, followed by community benchmarks, SGLang day-0 support, and quantized versions that run on as little as 6GB VRAM.

2026-09-18 ~ 2026-09-23 · 10 episodes · 68 posts

Episode 1 · Qwen Image 2.1 open-source model appears set for release (2026-09-18, 2 posts)

Reddit users spotted Qwen Image 2.1, a coming open-source image model from Alibaba, listed as Coming Soon. Anthropic researcher Andrew Carr had earlier teased another Qwen image generation model on the way.

Episode 2 · Early Qwen Image 2.1 hands-on reports split: fast and good at editing, weak photorealism (2026-09-19, 7 posts)

Ahead of its official launch, Qwen Image 2.1 was opened to selected creators, and hands-on reports surfaced on Reddit from Sept 19–21 with clearly split verdicts: some call it the strongest open-source image model to date, while others find it behind GPT Image 1.0 and weak on photorealism. Overall, speed, editing and reference-image capabilities are widely praised; realistic quality and artifacts are the main weaknesses, and one user found it can clean up GPT images' signature noisy texture.

Confirmed

  • Sanders (relayed by @Crazy-Repeat-2006), an early-access holder, compared it against mainstream open-source image models and concluded it is the strongest open-source image model so far, with realism exceeding ZIT and consistent output at high resolution; the Model S comparison was not fully detailed.
  • @13baaphumain generated 200+ "1girl" test images, judging the quality quite good for a 7B-parameter model with solid editing and reference-image skills, though with noticeable noise.
  • @AltruisticPrint717 tested in a ComfyUI workflow and was disappointed: quality below GPT Image 1.0, more errors and artifacts, like a "weakened" GPT Image 1.0, though faster.
  • @NoDaikon3851's first run used CFG 1, 40 steps, 2MP, Euler/Simple sampling and int8 quantization: artistic styles acceptable, photorealistic images largely poor.
  • @Choidonhyeon, trying it on Comfy Cloud, found that re-generating GPT-created images through Qwen Image 2.1 can somewhat remove their signature noisy texture.
  • A Reddit roundup by @fruesome collected output links from X creators including Hakm, Richard, Dev, Ivan and Ai Biceps, showcasing the version across styles and scenes.

Why it matters

  • This is a major iteration of the Qwen Image series; early word of mouth will shape community adoption after release.
  • The split centers on photorealism: art styles, editing and speed are praised, while realism and artifacts lag—useful for users choosing workloads.
  • The re-generation de-noising trick adds a practical role as a post-processing cleanup tool for GPT images.
  • If a 7B model approaches closed-source rivals, it sets a benchmark for the open-source image generation ecosystem.

Episode 3 · Alibaba's Qwen-Image-2.1 Announced as Fully Open-Source with Day-One ComfyUI Support (2026-09-20, 7 posts)

Alibaba's Tongyi team, together with ModelScope, officially announced that Qwen-Image-2.1, the latest generation of the Qwen-Image text-to-image model, will be open-sourced: after a 10-hour countdown, both weights and code would be released for download, local running, and downstream development. On the same day, a related PR was merged into ComfyUI and an official Comfy page went live, making ecosystem support nearly simultaneous with the release.

Confirmed

  • The Tongyi team and ModelScope's official accounts jointly teased the open-sourcing of Qwen-Image-2.1, with weights and code downloadable for self-deployment (m4, m5)
  • Reddit user CeFurkan posted a 10-hour countdown pointing to the public open-source release (m2, m3)
  • The related PR was merged into ComfyUI with a YouTube video showing various generation examples; Reddit user fruesome flagged the imminent release based on this (m1)
  • Comfy's official page comfy.org/qwen-image-2.1 is live, confirming upcoming ComfyUI workflow support (m7)

Timeline

  • 09-20: LiteBlackEagle surfaced the "Qwen/Qwen-Image-2.1" naming clue, sparking speculation (m6)
  • 09-20: fruesome found the merged ComfyUI PR and reported the release was coming the next day (m1)
  • 09-20: Tongyi and ModelScope officially announced the countdown; weights and code in 10 hours (m4, m5)
  • 09-20: Comfy's official page went live (m7)

Why it matters

  • Qwen-Image is one of the mainstream open-source text-to-image model families; version 2.1 ships with both weights and code, enabling direct local deployment and secondary development
  • Day-one ComfyUI integration and an official support page mean ecosystem adaptation is nearly synchronized with the release, so users can put the model to work in real workflows immediately

Episode 4 · Qwen Releases Prompt-Rewriting Models for Image Generation (2026-09-20, 2 posts)

Qwen released two fine-tuned Qwen3.5-VL 9B prompt-rewriting models for Qwen-Image-2.1, supporting both image-to-image and text-to-image prompt rewriting, with Qwen-Image-2.1-PE-T2I already available on Hugging Face.

Episode 5 · Qwen Image Model Beats Flux in Quality and Speed, Users Report (2026-09-20, 4 posts)

Community tests show Qwen Image 2.1 outperforms Flux1dev in quality and speed, runs on 8GB VRAM with int8 quantization, supports transparent backgrounds, and delivers strong reference-image consistency in generation and editing tasks.

Episode 6 · Alibaba Open-Sources Qwen-Image-2.1: One 7B Model for Image Generation and Editing (2026-09-20, 36 posts)

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

16 more related posts →

Episode 7 · Qwen Image 2.1 Tested: Strong Text Rendering, Yellowish Tones (2026-09-21, 2 posts)

Hands-on tests of Qwen Image 2.1 show it generates 1MP images in 25 seconds on RTX 5070/80 with strong text rendering, though results have a yellowish, GPT-Image-like tone and weaknesses in outfit swapping.

Episode 8 · Qwen-Image 2.1 Lands on SGLang with 2.55x Speedup from KV Cache (2026-09-21, 2 posts)

SGLang-Diffusion announced day-0 support for Alibaba's Qwen-Image 2.1, generating images in about 18.7 seconds on a single 4090, while a technical thread explains how the 7B DiT model's KV caching yields a 2.55x generation speedup.

Episode 9 · Qwen-Image-2.1 GGUF Quantized Version Tops Hugging Face Trending (2026-09-21, 4 posts)

A user-released GGUF quantized version of Qwen-Image-2.1, including an uncensored variant, has topped Hugging Face trending charts. It enables local text-to-image generation via ComfyUI, saving around 17GB of VRAM with little speed loss.

Episode 10 · Unsloth's quantized Qwen-Image-2.1 runs locally on just 6GB VRAM (2026-09-21, 2 posts)

Unsloth released FP8 and Dynamic GGUF quantized versions of Qwen-Image-2.1 on Hugging Face, enabling local image generation and editing on as little as 6GB of VRAM, claiming performance on par with Nano Banana 2.0.