How can I replace an object using a reference image while preserving its exact design?

A user reports that Qwen-Image-Edit-2511 struggles with one-to-one replacement of multiple identical objects in a scene, often removing all but one instance or misaligning viewpoints despite reference images. The user also seeks to reduce inference time from 6–7 minutes to 30–60 seconds on an NVIDIA A100-SXM4-80GB GPU, noting that torch.compile cuts subsequent runs to ~2 minutes, and asks if such speed is achievable without reducing the 40 inference steps.

Hi, first of all, a huge thanks to you Qwen gives me very good results with simple and medium-complexity images. However, I noticed one issue when there are multiple instances of the same object in a room. For example, if there are four sofas and I ask it to replace the sofas with chairs, its behavior is not very consistent. Sometimes, it removes all the sofas and places only one chair. If I explicitly prompt it to replace all the sofas and place a chair in each of the four positions, the viewpoints/angles of the generated chairs are often not correct, even when I provide 3–4 reference images of the chair. Overall, the performance seems very good, but this is the main limitation I have noticed so far: when there are multiple instances of the same object in an image, achieving accurate one-to-one replacement while preserving the original position and viewpoint is difficult. Here is my other issue: I’m trying to optimize the inference time. Current setup: GPU: NVIDIA A100-SXM4-80GB PyTorch: 2.13.0+cu130 CUDA: 13.0 Precision: BF16 Inference steps: 40 true cfg scale : 4.0 guidance scale : 1.0 Product-consistency LoRA: FractalAIResearch/Kalaido-qwenedit-lora optional VRAM usage: ~54–55 GB Currently, a single image takes around 6–7 minutes without optimization. After using torch.compile , the first generation takes around 6–7 minutes , but subsequent generations take around 2 minutes . My target is to get the inference time down to seconds ideally ~30–60 seconds while keeping 40 inference steps and similar output quality . Is this kind of inference speed achievable with Qwen-Image-Edit-2511 on an A100 80GB? Are there recommended optimizations, faster/quantized variants, or inference techniques that can bring the generation time into the seconds range without reducing the steps?