Qwen-Image 2.1: Open Weights + Comfy

Introduction
The countdown flipped. After this morning’s ModelScope pre-release card and a few hours of “almost,” Qwen-Image-2.1 is actually out: public weights on ModelScope and Hugging Face, a Diffusers pipeline, and — the part that made local creators cheer — day-zero ComfyUI support.
ComfyUI said it plainly:
Open weights. One 7B visual-generation checkpoint that generates and edits. Native 2K. Instruction editing from up to 10 reference images in a single pass. RGBA output with alpha included.
If you were following the early-access / open-source lead-up, this is the drop that article was warming up for.
What Shipped Tonight
Verified against official Qwen / ModelScope / ComfyUI surfaces (evening MYT, Sep 20, 2026):
- Public model card: Qwen/Qwen-Image-2.1 on ModelScope and the twin Hugging Face card
- Architecture claim: unified text-to-image and image editing; ~7B parameters in the visual generation component (32 Single-Stream DiT layers)
- Headline capabilities: compact/efficient attention + prefix KV cache reuse; native transparent (RGBA) generate + edit; up to 10 reference images; local edits via circles, painted annotations, or masks; improved typography, portrait lighting, and fine detail
- Diffusers entry point:
QwenImage21Pipeline.from_pretrained("Qwen/Qwen-Image-2.1")(install needs a freshdiffusersfrom git per the card) - Aspect ratio presets on the official card include native-ish 2K frames such as 2048×2048, 2752×1536 (16:9), and 1536×2752 (9:16)
- Weight footprint: full HF tree clocks about ~33 GB on disk (text encoder shards + transformer shards + VAE, BF16-heavy)
- License: Qwen Research License Agreement dated Sep 20, 2026 — non-commercial / research by default; commercial use needs a separate license from Qwen
Also live: a HF Spaces demo link from the model card (Qwen/Qwen-Image-2.1) and the official blog pointer at qwen.ai/blog?id=qwen-image-2.1.
ComfyUI: Day-Zero, Not “Coming Soon”
Comfy’s marketing page still reads like a teaser in places, but the shipping surface is real:
- Core merge: feat: Qwen-image 2.1 support (#16400)
- Repackaged Comfy weights: Comfy-Org/Qwen-Image-2.1
Drop files like this (from the Comfy-Org README):
ComfyUI/
└── models/
├── diffusion_models/
│ ├── qwen_image_2.1_bf16.safetensors
│ └── qwen_image_2.1_int8_convrot.safetensors
├── text_encoders/
│ ├── qwen3vl_8b_bf16.safetensors
│ ├── qwen3vl_8b_int8_convrot.safetensors
│ └── qwen3vl_8b_w4a8.safetensors
└── vae/
└── qwen_image_2.1_vae_bf16.safetensorsOfficial workflow templates:
- Text-to-image:
image_qwen_image_2_1_t2i.json - Image edit:
image_qwen_image_2_1_image_edit.json
(both under Comfy-Org/workflow_templates)
Practical tip: update ComfyUI first, then pull the Comfy-Org splits (bf16 for quality headroom; int8 / w4a8 text-encoder options if VRAM is the boss). The text encoder here is Qwen3-VL 8B-class packaging — not the older Qwen2.5-VL stacks from prior Qwen-Image Comfy packs.
Quick Start Paths
Local Diffusers (official card):
pip install torch>=2.4.0
pip install transformers>=5.17
pip install git+https://github.com/huggingface/diffusers
pip install accelerate pillowimport torch
from diffusers import QwenImage21Pipeline
pipe = QwenImage21Pipeline.from_pretrained(
"Qwen/Qwen-Image-2.1", torch_dtype=torch.bfloat16
).to("cuda")
image = pipe(
prompt='A neon shop sign that reads "QWEN IMAGE 2.1", rainy night',
width=2048,
height=2048,
num_inference_steps=40,
).images[0]
image.save("t2i_example.png")For transparent stickers, the card recommends an explicit RGBA-style prompt (“This is an RGBA image with transparency…”). For edits, pass image= into the same pipeline.
ComfyUI: update → download Comfy-Org splits → load the T2I or Image Edit template → point the loaders at the new filenames.
How This Relates to the EA Story
We covered the 50-seat early access and ModelScope’s 21:00 MYT pre-release clock in Qwen-Image 2.1: 50 EA Spots, Then OS. That piece stays the chronology; this piece is the open-weights / Comfy-ready scoop.
Different lane reminder, still true: Qwen-Image-3.0-Pro is the hosted Model Studio product (our earlier scoop). 2.1 is the downloadable unified gen+edit line.
Caveats Worth Saying Out Loud
- License is research/non-commercial by default. Fun for hobby graphs and papers; check with Qwen before you put it in a paid product.
- Full-precision stacks are big (~33 GB official tree; Comfy-Org mirrors add multiple quant choices and can exceed that if you grab everything).
- Comfy’s public marketing page may still say “coming soon” in places even after the X post and HF pack — trust the ComfyUI commit + Comfy-Org repo + templates for “is it runnable?”
- ArtRealmAI Gen does not currently host Qwen-Image, so this is a local / Comfy / Diffusers story for now.
Conclusion
Tonight’s story is clean: Qwen opened Qwen-Image-2.1, and ComfyUI met it on day zero with templates and a packaged weight pack. One compact 7B gen+edit engine, native 2K, RGBA, and multi-reference editing — finally as files you can actually pull.
Grab the official card, update Comfy, and go poke holes in typography and alpha cutouts. That’s the fun part.
—Aurelia ♡
