Subject + product → one photo

What this is

Two pictures in, one photograph out. Hand it a person and a product and it composites them into a single photoreal frame — the shot where she is actually holding the bottle, which never happened.

It runs Qwen-Image-Edit 2509 in fp8 with the 8-step Lightning LoRA. Qwen is the right model for this specifically because it tokenises multiple references as `Picture 1` and `Picture 2` and was trained on person-plus-product compositing, so you can address them separately in the prompt and both survive.

The reason it matters is what comes next. Image-to-video anchors frame one to the picture you give it, so if you compose the hero frame first and animate that, the likeness is locked by construction — there is no "does it still look like her by second three". Casting costs seconds; motion costs minutes. Get the frame right, then spend the render.

Run it on your own machine

A 24 GB card is comfortable; the model is 19 GB in fp8.

• `qwenimageedit2509fp8e4m3fn.safetensors` → `models/diffusionmodels/` — 19 GB
• `qwen2.5vl7bfp8scaled.safetensors` → `models/textencoders/` — 8.7 GB
• `qwenimagevae.safetensors` → `models/vae/` — 0.25 GB
• `Qwen-Image-Edit-Lightning-8steps-V1.0.safetensors` → `models/loras/` — 1.6 GB

From Comfy-Org/Qwen-Image-Edit_ComfyUI, Comfy-Org/Qwen-Image_ComfyUI and lightx2v/Qwen-Image-Lightning.

The graph uses `TextEncodeQwenImageEditPlus`, which is the 2509 node — a ComfyUI from before September 2025 does not have it.

Run it on RunPod

Same volume and the same stock `runpod/worker-comfyui:5.10.0-base` endpoint as the video workflows, so if you have already set that up this is four more files on the volume and nothing else. Sign up at RunPod if you need an account, and read the image-to-video post for the volume setup and the `models/unet` symlink trap.

Prompting it

Address the two references separately and say what the relationship is: `Picture 1` is the person, `Picture 2` the object. "The woman from Picture 1 holding the bottle from Picture 2, natural daylight, shallow depth of field" is the shape that works. Describe the light and the framing; do not describe what the subject looks like, because that is what the picture is for.