Qwen Image Edit 2511 Angle Camera

Re-shooting an existing image from a new camera angle used to be the ugliest job in the workflow: extract an OpenPose skeleton, wire up a ControlNet checkpoint, and still fight the pose. In late 2025 that whole pipeline got replaced by a LoRA on an edit model. This node is the ComfyUI wrapper for the specific LoRA everyone ended up using - the Qwen-Image-Edit-2511 Multiple-Angles LoRA - and it turns "change the camera" into three sliders and a <sks> prompt you never have to type.

Here's the one-line version: it loads the fixed angle LoRA onto your Qwen-Edit model and CLIP, and hands you back the exact camera prompt that LoRA was trained to respond to. You drag, it conditions, you run.

How it works

Qwen-Image-Edit-2511 is the current open weight of Alibaba's instruction editor (Apache 2.0, and the license is the reason its LoRA ecosystem exists at all). The Multiple-Angles LoRA is trained on the editor - it does nothing on the base Qwen-Image text-to-image model, which is a distinction that bites people. This node hardcodes that one LoRA and builds the trigger prompt for you.

Three controls map onto the camera vocabulary the LoRA learned:

• azimuth (0–315) snaps to eight compass directions: front, front-right, right side, and so on around the subject.
• elevation (-30 to 60) picks low-angle, eye-level, elevated, or high-angle shot.
• distance (0.6–1.8) chooses close-up, medium, or wide shot.

Those get combined into a prompt like <sks> front-right quarter view eye-level shot medium shot. In the browser UI there's a 3D camera map rendered inside the node - drag to move azimuth and elevation, scroll the mouse wheel for distance - which beats typing numbers if you're thinking spatially. The strength_model and strength_clip sliders are the LoRA weights (1.0 is a sensible start); set both to 0 and the node silently skips the LoRA and emits an empty prompt, which is handy when you want the model back untouched.

The LoRA itself is not in the repo, on purpose. auto_download_lora (on by default) fetches qwen-image-edit-2511-multiple-angles-lora.safetensors from fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA on Hugging Face the first time it's needed and drops it in your first configured loras folder. Turn it off and it expects the file at ComfyUI/models/loras/qwen-image-edit-2511-multiple-angles-lora.safetensors - if it's not there, you get a clear FileNotFoundError instead of a silent no-op.

The inputs and outputs that matter

Required inputs are model, clip, image, and the three camera sliders. Note the image input: the Python side doesn't actually touch it (the LoRA math is all model + CLIP), so it's really there to keep your graph honest - wire your source image in anyway. The outputs are the conditioned model and clip, plus the prompt string. Feed model and clip straight into the sampler path of a standard Qwen-Edit workflow and route the prompt into your CLIP Text Encode. This node does the conditioning; the actual edit happens in the normal KSampler steps around it.

Installing it

It's one node in the four-node ComfyUI-Qwen-FrameKit pack:

cd ComfyUI\custom_nodes
git clone https://github.com/uron83/ComfyUI-Qwen-FrameKit.git

or search "Qwen FrameKit" in ComfyUI Manager. Restart ComfyUI and hard-refresh the browser. The pack's requirements.txt is literally "No external Python dependencies." - the real download is the LoRA (a few hundred MB), and the bigger one is the 20B Qwen-Image-Edit-2511 checkpoint itself. Run the GGUF quantizations unless you have a lot of VRAM; that's what makes this workflow affordable.

Where people get burned

• Left/right confusion. It's real and it's the LoRA, not you. Community reports of the Multiple-Angles family consistently note it mixing up sides - a "right side view" comes back as a mirror. Generate a short angle sequence and eyeball it rather than trusting the label.
• Input resolution matters. The community sweet spot is roughly a 2 megapixel input; tiny images give the angle control little to work with. Size your source before the edit (SmartAspectResize in the same pack is built for exactly this).
• Wrong revision. This node is pinned to the 2511 LoRA. If you're on Edit-2509, the 2509 Multiple-Angles LoRA is a different file and this node won't find it.
• It re-emits the whole frame. Qwen-Edit returns a new image, so untouched pixels drift and faces can wander, worst on faces the model hasn't seen. For pixel-exact edits, crop to a mask, edit the crop, stitch it back.