Qwen-Image¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
qwen-image |
Qwen/Qwen-Image | 75e0b4b |
Apache-2.0 | Verified |
qwen-image-2.1 |
Qwen/Qwen-Image-2.1 | d26bb61 |
Qwen Research License | Verified |
qwen-image-edit (+ image input) |
Qwen/Qwen-Image-Edit | ac7f931 |
Apache-2.0 | Verified |
qwen-image-edit-2509 (+ image input) |
Qwen/Qwen-Image-Edit-2509 | d3968ef |
Apache-2.0 | Verified |
Catalog: qwen-image-2 ·
Models: qwen-image, qwen-image-2.1, qwen-image-edit, qwen-image-edit-2509 · Task: t2i
Architecture. The original release uses Qwen2.5-VL, a 60-layer patch-2 dual-stream transformer and a 16-channel causal Qwen VAE. Qwen-Image 2.1 uses Qwen3-VL-8B, a single-stream transformer with block-causal joint attention and a text KV cache, and a 64-channel causal VAE. Both releases use dynamic-shift flow-match Euler sampling and have a native image size of 1024×1024.
Usage.
from imagehub import pipeline
with pipeline(
"t2i", backend="qwen-image-2.1", model="checkpoints/qwen-image-2.1", device="mps"
) as pipe:
result = pipe("a cat, oil painting", seed=42, negative_prompt="blurry", true_cfg=4.0)
result.images[0].save("cat.png")
Use qwen-image in both commands for the original release. Its steps and true_cfg defaults are
50 and 4.0, and dimensions must be multiples of 16. For Qwen-Image 2.1, steps defaults to 40,
true_cfg defaults to 1.0, and dimensions must be multiples of 32. Both accept negative_prompt,
num_images, seed and latents; unrecognized settings raise an error.
Seed compatibility. On CPU, the same seed and settings give the same image as the pinned
diffusers reference with generator=torch.Generator("cpu").manual_seed(seed)
(validation). MPS and CUDA runs are smoke tests only; there is no claim of exact
equality on these devices.
For instruction editing, download qwen-image-edit (one input image) or
qwen-image-edit-2509 (one or more input images), then pass an imagehub.Image or Pillow image:
with pipeline("t2i", backend="qwen-image-edit", model="checkpoints/qwen-image-edit") as pipe:
result = pipe("replace the sky with a sunset", image=source, seed=42)
The edit releases use the Qwen2.5-VL vision tower to condition the instruction and append VAE
latents from each source image to the denoiser input. Their defaults are 50 steps and true_cfg=4;
dimensions must be multiples of 16. The real-checkpoint reference tests intentionally use two
steps at 32×32 because the published transformer and conditioner are large.
Transformer LoRA adapters (lora=, lora_scale=, lora_fuse=) load on the edit releases as on
the other Qwen-Image releases and match diffusers' LoRA loader in the tiny tests.
Out of scope. Video inputs and fp16/bf16 parity are not supported.
The original release's published bfloat16 weights are upcast for CPU float32 validation. Its real-checkpoint run took 30 minutes on an M2 Max (CPU) and every recorded comparison is exactly 0 (validation).
License. Original Qwen-Image weights use Apache-2.0. Qwen-Image 2.1 weights use the Qwen Research License and are for research use; see each repository's license. imagehub does not distribute weights.