import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("wavespeed/Qwen-Image-bf16", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]Qwen-Image-bf16
An unmodified bf16 copy of Qwen/Qwen-Image, resharded for WaveSpeed's loader.
What was changed
Nothing numerically. Tensor names, shapes, dtypes (bf16) and total parameter bytes are identical to upstream; the transformer is simply written as 5 shards instead of 9, which suits how our serving nodes stream weights from cache.
If you are not us, use Qwen/Qwen-Image
instead. This repository exists as a pinned serving artifact and offers no
advantage over the original.
The quantized variants derived from this baseline are
wavespeed/Qwen-Image-e4m3
(fp8) — those do differ from upstream.
License
Apache-2.0, inherited from Qwen-Image.
- Downloads last month
- 15