berkerdooo/qwen-image-edit-2511-camera-angle-lora

Qwen-Image-Edit-2511 Camera-Angle LoRA

Give one image of an object and a camera move in degrees; the model renders the same object from the new angle.

Prompt format (numbers can be any value):

Move the camera 135 degrees to the right around the object. Target camera elevation: 20 degrees. Keep the object, its colors, and the white background the same.

• Azimuth: 0-180 degrees, "right" or "left" (the camera orbits around the object).
• Elevation: about -10 to 65 degrees.

How to use

import torch
from diffusers import QwenImageEditPlusPipeline
from diffusers.utils import load_image

pipe = QwenImageEditPlusPipeline.from_pretrained("Qwen/Qwen-Image-Edit-2511", torch_dtype=torch.bfloat16).to("cuda")
pipe.load_lora_weights("berkerdooo/qwen-image-edit-2511-camera-angle-lora")

image = load_image("object_on_white.png")   # one object, centered, plain background
prompt = ("Move the camera 120 degrees to the right around the object. Target camera elevation: 20 degrees. "
          "Keep the object, its colors, and the white background the same.")
out = pipe(image=[image], prompt=prompt, num_inference_steps=25, true_cfg_scale=3.0,
           generator=torch.manual_seed(0)).images[0]
out.save("new_view.png")

Training

• Base: Qwen/Qwen-Image-Edit-2511, LoRA rank 32, 8,000 steps, batch 2, 512 px, AdamW8bit, lr 1e-4 (ai-toolkit).
• Data: 4,899 (source view, target view) pairs from 1,633 Objaverse objects (plus 40 held-out test objects) with CC-BY 4.0 or CC0 licenses, rendered by Cap3D (20 random views per object, white background). Rooms, panoramas, buildings and maps were filtered out.
• Hardware: 1x RTX PRO 6000 (96 GB), 15 h 43 min.
• Attribution for every source object: ATTRIBUTION.csv.

Limitations

• Trained on synthetic 512 px renders with flat lighting and white backgrounds; real photos with complex backgrounds may need background removal first.
• Each output is a separate generation: views are not guaranteed to be 3D-consistent.
• Large rotations of thin or flat objects are the weakest case.
• Evaluated only by visual comparison on held-out objects so far.

Disclaimer

Independent fine-tune, not affiliated with Alibaba / the Qwen team. Provided as is under Apache 2.0, without warranty.