FireRed Loader (Trent)

The first thing this node does is download roughly 40GB onto your machine, so before you click it, know what you're in for. FireRed Loader (Trent) is the entry point to FireRed-Image-Edit-1.0, an Apache-2.0 instruction editor from the FireRed team - and, as the community established within days of the February 2026 release by comparing layer similarity, it's essentially Qwen-Image-Edit-2509 under a different name (~99.96% of the same weights, a shallower change than 2509 → 2511). That's not a criticism, it's context: it means everything Qwen-Edit became famous for - "change her dress to blue", remove the watermark, edit the text inside an image - is exactly what this loader sets up.

It's the "downloads it, loads it, caches it" node of the pack. It pulls the model through the HuggingFace snapshot_download helper the first time you run it, stores it in ComfyUI/models/diffusers/FireRedTeam--FireRed-Image-Edit-1.0, then loads it via diffusers' QwenImageEditPlusPipeline. And it caches the loaded pipeline in memory, so the second run doesn't re-parse those tens of gigabytes - it just hands you the pipe.

The three inputs that actually matter

• model_id - the HuggingFace repo ID or a local path, default FireRedTeam/FireRed-Image-Edit-1.0. If you already downloaded the repo elsewhere, point it at the folder and it skips the download.
• precision - bf16 or fp16, default bf16. bf16 is the right call on any Ampere-or-newer card; reach for fp16 only if bf16 misbehaves on your hardware.
• offload - model_cpu_offload (default), sequential_cpu_offload, or full_gpu. This is your VRAM budget dial. Keep the default unless you know you've got the headroom; full_gpu loads everything onto the card and will OOM a consumer GPU.

The single output is firered_pipe (type FIRERED_PIPE) - wire it into the FireRed Sampler, the LoRA loader, or the Unloader. It's not a checkpoint in the usual ComfyUI sense; it's a live diffusers pipeline object, which is why it needs the pack's own nodes on the other end.

Install

ComfyUI Manager (search "ComfyUI-FireRedTrent"), or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/ComfyUI-FireRedTrent
# restart ComfyUI - Manager or comfy-cli will pull requirements.txt

The pack's requirements.txt pins diffusers>=0.36.0, transformers>=4.45.0, accelerate, huggingface_hub, and peft. The one that actually bites is diffusers - the loader refuses to start unless your install has QwenImageEditPlusPipeline, and it prints the exact pip install diffusers>=0.36.0 command if it doesn't.

Where people get burned

The download, first. This is a 20B-class model at bf16 - tens of gigabytes - and the first load takes a while even on fast storage, with the progress bar only covering the load, not the download. Second: VRAM. This pack loads the full diffusers pipeline, not a GGUF quant, so on a 12–16GB card you're choosing between model_cpu_offload (works, but slow - the CPU↔GPU shuffling adds up over 40 steps) and OOM. The community standard for this model family on consumer hardware is quantized weights, which this pack doesn't do. Third: the cache cuts both ways. The pipeline stays in memory between runs - great for iteration, exactly why the Unloader exists. If you're coming from Comfy-Org's official Qwen-Image-Edit nodes, the same model runs there too; this loader earns its place when you want the whole FireRed pack in one install.