Qwen-Image-Edit-2511 takes an existing image plus an instruction and edits it, keeping identity, clothing, and lighting intact. We run it two ways: re-angling a still to a new camera viewpoint (a small LoRA constrains it to camera-angle language only — azimuth, elevation, distance), and free-form pose/position edits from a plain-English instruction with no LoRA at all.
What it needs: ~12 GB VRAM. Runs a 4-step Lightning schedule (steps/cfg/sampler are fixed, not user-tunable) for fast turnaround.
What it's good for: getting a second angle of a shot you already like without re-rolling the whole generation, or changing what a subject is doing while everything else in the frame stays put.
In FlixML: `qwen_multiangle` (re-angle via a constrained camera-angle LoRA) and `qwen_pose_edit` (free-form instruction edit, no LoRA).