Qwen-Image-2.1 LoRA Lets Developers Rotate Any Subject With Camera Prompts
A new camera-control adapter for Qwen-Image-2.1 lets you reshoot any photo from 72 preset framings using a simple prompt grammar.
Takeaways
- First camera-control LoRA for Qwen-Image-2.1, released under Apache-2.0.
- Supports 72 framings via a tiny
<mva>prompt grammar covering 12 azimuths and 4 elevations. - v2 doubles rank to 64, uses bf16, trained on 13,328 pairs; step 1,500 recommended.
- Works with diffusers, ComfyUI, and a Gradio Space at strength 0.8 to 1.0.
- Generalizes to real photos; fine fur and identity detail soften on OOD subjects.
- Author flags that CLIP cosine benchmarks fail to measure actual pose change.
Rotating a subject in an AI-generated image often requires repeated prompts that change its identity or composition. A new Multiple-Angles LoRA for Qwen-Image-2.1 adds explicit camera control: provide a reference image, choose an azimuth and elevation, and generate the subject from that viewpoint. The base model handles generation and editing in one pipeline and supports RGBA output, including alpha transparency.
A LoRA is a compact set of adapter weights loaded alongside a base model. This 319 MB adapter changes how Qwen-Image-2.1 interprets viewpoint prompts while attempting to preserve the reference subject and scene.
| Item | Details |
|---|---|
| Recommended release | V2, rank 64, checkpoint step 1,500 |
| Input | Reference image plus named camera labels |
| Training resolution | 512 × 512 pixels |
| Interfaces | Diffusers, ComfyUI, and Gradio |
| License | Apache-2.0 |
Camera control through fixed labels
The adapter treats camera direction as a controlled prompt vocabulary. Every prompt begins with <mva>, followed by an azimuth label for the horizontal position around the subject and an elevation label for the camera height. Examples include back-left quarter view, eye-level shot, high-angle shot, and the optional close-up modifier.
The model card reports 36 base poses and 72 framings with close-ups. It describes 12 azimuths at 30-degree intervals crossed with four elevations, which produces 48 base combinations before adding close-up variants. Those counts do not reconcile, so integrations should expose combinations explicitly documented by the author; support for the full Cartesian product remains unclear.
Minimal diffusers call
# pipe is a QwenImage21Pipeline in edit mode with the v2 LoRA loaded
image = pipe(
prompt="<mva> back view, eye-level shot",
image=input_image,
num_inference_steps=40,
true_cfg_scale=1.0,
).images[0]