Skip to main content

Qwen-Image (Open-Weight)

Qwen-Image is Alibaba's open-weight image foundation family on Hugging Face — the open counterpart to the hosted Qwen-Image DashScope API. Use this for text-to-image, image editing, and virtual try-on on your own GPU (privacy, no per-request cost, offline).

Apache-2.0 weights. OpenTryOn uses Hugging Face Diffusers (not ComfyUI).

Hosted twin: opentryon generate|edit|vton --model qwen-image.

Capabilities​

TaskCLIDefault weightsPipeline
Generate (T2I)--model qwen-image-localQwen/Qwen-Image-2512QwenImagePipeline
Edit / I2I (1–3 refs)--model qwen-image-localQwen/Qwen-Image-Edit-2511QwenImageEditPlusPipeline
VTON--model qwen-image-localsame Edit-Plusperson + garment (up to 3 refs)

Earlier open checkpoints still work via env overrides:

RoleOther official ids
T2IQwen/Qwen-Image (original 20B)
Single-image editQwen/Qwen-Image-Edit (QwenImageEditPipeline, 1 ref only)
Multi-image editQwen/Qwen-Image-Edit-2509

Qwen-Image 2.0 / 3.0 are the hosted DashScope ids (qwen-image-2.0, qwen-image-3.0-pro). There are no matching open Diffusers weights for those hosted SKUs yet — local OpenTryOn tracks the official HF open series (2512 T2I, 2511 Edit).

Overview​

FeatureValue
LicenseApache 2.0
Architecture~20B MMDiT + Qwen2.5-VL text encoder
Packed sizesOfficial aspect map, e.g. 1:1 → 1328×1328
Min VRAM~40GB+ bf16 full load; cpu_offload on by default for smaller cards
OpenTryOn idqwen-image-local (extra="local")

VTON is person + product composition (Edit-Plus was trained on combinations like person+product / person+scene), not a dedicated garment-fit model.

Installation​

pip install opentryon[local]
pip install -U diffusers transformers accelerate
# Qwen-Image needs transformers>=4.51.3 (Qwen2.5-VL) and a Diffusers
# build that ships QwenImagePipeline / QwenImageEditPlusPipeline.

CUDA GPU required. First run downloads tens of GB of weights.

Environment​

# Optional overrides (defaults shown)
# QWEN_IMAGE_LOCAL_MODEL_ID=Qwen/Qwen-Image-2512
# QWEN_IMAGE_EDIT_MODEL_ID=Qwen/Qwen-Image-Edit-2511
# QWEN_IMAGE_LOCAL_PATH=/path/to/local/t2i-snapshot
# QWEN_IMAGE_EDIT_PATH=/path/to/local/edit-snapshot
# HF_TOKEN=...   # only if the Hub repo is gated for your account

CLI​

opentryon generate --model qwen-image-local \
  --prompt "editorial lookbook, linen trench on a sunlit terrace" \
  --aspect-ratio 16:9 --seed 42
# Optional: --model-id Qwen/Qwen-Image  (original 20B T2I)

opentryon edit --model qwen-image-local \
  --images person.jpg \
  --prompt "Change the jacket to black leather, keep the face and pose."

opentryon vton --model qwen-image-local \
  --person-image model.jpg --garment-image garment.jpg \
  --garment-description "olive green bomber jacket"

Pair with local or hosted understand:

opentryon understand --model qwen3.8 --image garment.jpg \
  --prompt "Describe this garment: category, color, fabric, cut."

MCP​

  • generate_qwen_image_local
  • edit_qwen_image_local
  • vton_qwen_image_local

Requires pip install opentryon[local] on the machine running the MCP server.

Python​

from tryon.models import QwenImageLocalAdapter

adapter = QwenImageLocalAdapter()  # downloads 2512 / 2511 on first use

images = adapter.generate_text_to_image(
    "editorial lookbook, linen trench on a sunlit terrace",
    aspect_ratio="16:9",
    seed=42,
)

edited = adapter.generate_image_edit(
    "person.jpg",
    prompt="Change the jacket to black leather, keep the face and pose.",
)

tryon = adapter.generate_virtual_tryon(
    person="model.jpg",
    garment="garment.jpg",
    garment_description="olive green bomber jacket",
)
tryon[0].save("qwen_tryon_local.png")

Use original T2I weights and Edit-Plus 2509:

adapter = QwenImageLocalAdapter(
    model_id="Qwen/Qwen-Image",
    edit_model_id="Qwen/Qwen-Image-Edit-2509",
)

QwenImageLocalAdapter​

class QwenImageLocalAdapter:
    def __init__(
        self,
        model_id: Optional[str] = None,       # T2I HF id / path
        edit_model_id: Optional[str] = None,  # Edit HF id / path
        device: Optional[str] = None,         # cuda required
        dtype: str = "bfloat16",
        cpu_offload: bool = True,
    )

Methods​

  • generate_text_to_image(prompt, aspect_ratio="1:1", width=None, height=None, num_inference_steps=50, true_cfg_scale=4.0, ...)
  • generate_image_edit(image, prompt, num_inference_steps=40, ...)
  • generate_multi_image(images, prompt, ...) — 1–3 refs on Edit-Plus
  • generate_virtual_tryon(person, garment, prompt=None, garment_description=None, ...)

Pipelines load lazily (T2I vs edit) so a generate-only run does not download the edit weights.

Hardware notes​

  • Full bf16 load is datacenter-class (~40GB+). cpu_offload=True (default) trades speed for VRAM.
  • Community FP8 / Nunchaku / Lightning LoRA stacks exist; OpenTryOn stays on official Diffusers pipelines. Pass a local quantized snapshot via model_id= / edit_model_id= if you already have one that loads with the same pipeline classes.
  • Do not expect Qwen-Image-2.0/3.0 DashScope quality from the 2512 open T2I checkpoint — they are different hosted SKUs.
  • Qwen/Qwen-Image-Layered (RGBA layer decomposition) is a separate Diffusers pipeline and is not registered in OpenTryOn yet. Pass a custom snapshot only if it loads with QwenImagePipeline / QwenImageEditPlusPipeline.

References​