---

name: inference description: "Run ICEdit single-image editing from the CLI, including normal LoRA inference, MoE inference, prompt templating, image-width normalization, output saving, seed control, and CPU-offload guidance." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Inference

Use this sub-skill for one-image, prompt-driven editing in the command line.

• Normal LoRA editing and MoE LoRA editing through one bundled helper.
• Normal mode is standalone; default model ids resolve through the Hub and weights are not bundled.
• The fixed diptych prompt template the helper prepends automatically.
• 512-pixel width normalization, height rescaling, and output naming.
• Model and LoRA selection, seed control, and --enable-model-cpu-offload.
• MoE mode requires an ICEdit checkout with vendored icedit/, passed via --repo-root.

What it does not cover

• browser or Gradio launch details; use the root ic-edit skill and its gradio route
• training launch, configs, and dataset prep; use the root ic-edit skill and its training route

Bundled files

• scripts/run_icedit_inference.py: one helper for both --mode normal and --mode moe
• references/quickstart.md: command templates and mode selection
• references/examples.md: bundled example images and width-normalization notes
• references/troubleshooting.md: common failure modes and fixes

Fast route

1. Pick --mode normal for standalone standard LoRA inference or --mode moe only with a checkout that contains icedit/.
2. Set --image and --instruction; the helper prepends the fixed prompt template for you.
3. Override --flux-path and --lora-path when you want local weights instead of Hub ids.
4. Add --enable-model-cpu-offload on lower-VRAM CUDA machines.
5. Read references/troubleshooting.md if CUDA, download, resize, or seed behavior looks off.