import torch
from diffusers import QwenImageEditPlusPipeline, QwenImageTransformer2DModel
from transformers import BitsAndBytesConfig
from PIL import Image

quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

transformer = QwenImageTransformer2DModel.from_pretrained(
    "mash2005/Qwen-Image-Edit-2511-MultiAngles-Q4",
    subfolder="transformer",
    quantization_config=quant_config,
    torch_dtype=torch.bfloat16,
    device_map="cuda"
)

pipeline = QwenImageEditPlusPipeline.from_pretrained(
    "mash2005/Qwen-Image-Edit-2511-MultiAngles-Q4",
    transformer=transformer,
    torch_dtype=torch.bfloat16,
)
pipeline.enable_model_cpu_offload()

image = Image.open("your_image.png").convert("RGB")

result = pipeline(
    image=[image],
    prompt="Show this object from the back view, azimuth 180 degrees",
    negative_prompt=" ",
    num_inference_steps=30,
    guidance_scale=1.0,
    true_cfg_scale=4.0,
).images[0]

result.save("output.png")

Prompt format

Use azimuth/elevation language for best results:

"front view, azimuth 0 degrees"
"back view, azimuth 180 degrees"
"side view, azimuth 90 degrees"
"top-down view, elevation 90 degrees"

Credits

Base model: Qwen/Qwen-Image-Edit-2511
LoRA: fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA