Qwen-Image-2.1-UltraFast
8-step image editing based on Qwen-Image-2.1. Built with Qwen · Non-commercial research.
Detailed documentation & benchmarks · Setup & usage · GGUF download
Image comparisons
Input → original model (40 steps) → UltraFast (8 steps). These are two recorded BF16 training examples. Each detail view is a crop of the same example.
Expression edit
“What if she had a sad face?”
Add a flower sticker
“Let's add a flower sticker on the fridge.”
Original image files · Image credits
At a glance
| Original | UltraFast | |
|---|---|---|
| Denoising steps | 40 | 8 |
| Recorded mean denoising time | 69.23 s | 17.89 s |
Historical comparison on two training cases, with loading, encoding and VAE decoding excluded; T40 includes recording callbacks. Quality retention is unmeasured. Full Flint charts and limitations.
Code: load the adapter
Use the included custom loader in a CUDA/BF16 environment. Installation instructions.
from pathlib import Path
import sys
import torch
from huggingface_hub import snapshot_download
from diffusers import QwenImage21Transformer2DModel
local = Path(snapshot_download(
"Haverbex/Qwen-Image-2.1-UltraFast",
allow_patterns=["adapter/*", "ultraedit_adapter.py", "configs/*", "LICENSE", "NOTICE"],
))
sys.path.insert(0, str(local))
from ultraedit_adapter import inject_dit_lora, load_adapter
transformer = QwenImage21Transformer2DModel.from_pretrained(
"Qwen/Qwen-Image-2.1",
subfolder="transformer",
revision="790c92633540aa0cb11d9abf19eb46d861714758",
torch_dtype=torch.bfloat16,
device_map={"": "cuda:0"},
)
adapters = inject_dit_lora(
transformer, rank=32, alpha=32, adapter_dtype=torch.bfloat16,
)
load_adapter(adapters, local / "adapter")
transformer.eval()
Download the loading script. This loads weights only. The full generation components and recorded eight-step schedule are explained in the usage guide.
Formats: BF16 adapter · Merged Q4_K GGUF, 3.77 GiB. GGUF image quality and speed have not been benchmarked.