FireRed Image Edit: ComfyUI setup and editing tests

This walkthrough introduces FireRed Image Edit 1.0, a model based on Qwen Image Edit, through a ComfyUI workflow and several editing tests. The presenter has not completed an extensive comparison with Qwen, so this is an initial assessment.

The setup covers BF16, FP8 and quantized GGUF downloads. The presenter suggests GGUF for low VRAM setups and recommends FP8 when memory permits, but gives no minimum hardware requirement. The workflow uses the Qwen Image Edit 2511 Lightning four-step LoRA, a Qwen2.5-VL 7B text encoder and Qwen Image VAE. The presenter describes sampling in four steps instead of the usual 20, with a limited quality tradeoff.

Controls select the model format and either a custom output size or the first input image's resolution. Up to three input images can be enabled independently. Tests cover old-photo restoration and colorization, transferring a tiger pattern onto clothing, and adding a hat from a separate reference.

A three-person cafe composition uses a custom 1280x720 output; the presenter notes some identity drift. A final single-image test changes a jacket to red, which the presenter says preserves the surrounding scene. Prompting can require adjustment, and a direct comparison with Qwen Image Edit is left for a later video.