One LoRA edits images better than GPT-4o

What it does

ICEdit is a NeurIPS 2025 method for instructional image editing built on Flux.1-fill-dev. Instead of fine-tuning the full diffusion transformer, it injects a small LoRA—trained on 0.5% of the data and 1% of the parameters that prior SOTA methods required—to perform multi-turn and single-turn edits from natural language instructions. It runs as a Hugging Face demo, a Gradio app, or inside ComfyUI workflows.

The interesting bit

The trick is framing edits as in-context generation: the model receives a diptych-style prompt describing the original scene and the desired change, and the LoRA learns to render the edit while preserving identity. A Mixture-of-Experts variant is available, and community ports have squeezed it down to 4 GB VRAM via quantization and NPU offloading.

Key highlights

• Claims to surpass GPT-4o on ID persistence, trained on just 0.5% of the data used by previous SOTA approaches
• Two LoRA variants: a normal LoRA for ComfyUI compatibility and a newer MoE-LoRA released in September
• Memory footprint spans 35 GB for native inference down to 4 GB with ComfyUI-nunchaku or 10 GB with GGUF quantization
• Supports multi-turn conversational editing; the README shows chains of precise sequential changes
• Training code is open-sourced, so you can bake your own editing LoRAs

Caveats

• Input images are rigidly resized to 512 px width; anything else gets squashed automatically
• The ComfyUI workflow requires a fixed pre-prompt format (“A diptych with two side-by-side images…”) and cannot load the MoE-LoRA with the standard ComfyUI LoRA loader
• The authors note that failed generations are common enough that they explicitly recommend retrying with a different seed

Verdict

Worth a look if you want local, instruction-driven image editing without renting a server farm. Skip it if you need arbitrary resolutions or a fully polished, seed-stable production pipeline.