Qwen-Image-Edit-2511 workflow consumes almost all 32GB VRAM and may run out of memory

[paulhe666]

I am using the qwen_image_edit_plus_2511 workflow in ComfyUI with ai-toolkit-inference.
When running Qwen-Image-Edit-2511 with a local LoRA, the VRAM usage becomes extremely high and almost fills my 32GB GPU memory.

In my case, VRAM usage can reach around 30GB+ / 32GB, and the generation becomes very slow or may run into CUDA out-of-memory issues.

Environment

• OS: Windows
• App: ComfyUI Desktop App
• Custom node: ai-toolkit-inference
• Pipeline: qwen_image_edit_plus_2511
• Model: Qwen/Qwen-Image-Edit-2511
• GPU: NVIDIA RTX 5090, 32GB VRAM
• LoRA: local .safetensors LoRA
• Offload mode: model

Steps to reproduce

1. Install ai-toolkit-inference in ComfyUI.
2. Load the Qwen Image Edit Plus 2511 workflow.
3. Load a local LoRA.
4. Run image generation with Qwen-Image-Edit-2511.
5. Observe GPU VRAM usage during model loading and generation.

Expected behavior

The workflow should be able to run more reliably on a 32GB VRAM GPU, ideally with lower peak VRAM usage or a recommended low-VRAM configuration.

Actual behavior

The workflow consumes almost all available VRAM on a 32GB GPU.
VRAM usage can reach around 30GB+ / 32GB, leaving very little margin. The process may become extremely slow or fail with CUDA out-of-memory.

Additional notes

I would like to know whether this VRAM usage is expected for Qwen-Image-Edit-2511, or if there are recommended settings to reduce memory usage, such as:

• lower-memory loading mode
• stronger CPU offload
• sequential CPU offload
• quantized loading
• lower precision loading
• lower resolution / steps
• LoRA loading method changes
• any recommended configuration for 24GB / 32GB GPUs

[kalvinrv]

Hi, thanks for the detailed report.

Yes, this VRAM usage is expected for Qwen/Qwen-Image-Edit-2511 with the current qwen_image_edit_plus_2511 pipeline, especially when running with a LoRA.

This is a very large bf16 Diffusers pipeline. The checkpoint includes a large text encoder, transformer, and VAE, and the model weights are much larger than 32GB on disk. With offload_mode=model, Diffusers moves major modules between CPU and GPU, but peak VRAM can still get very close to the full 32GB during loading/generation. Since this pipeline also fuses the LoRA weights via fuse_lora(), it can add additional temporary memory pressure.

For 32GB GPUs, we recommend trying:

• Set offload_mode to sequential for the lowest VRAM usage. This will be slower, but should reduce peak memory.
• Reduce output resolution if possible.
• Use fewer/lower-resolution reference images when possible.
• Close other GPU-heavy applications before running ComfyUI.
• If available in your setup, use a quantized/FP8 version of the model for more reliable low-VRAM inference.

For 24GB GPUs, the current bf16 pipeline is generally not recommended unless using very aggressive offload or a quantized model.

We’ll also consider documenting a recommended low-VRAM configuration for this workflow and looking into whether a quantized loading path can be supported in this custom node.