This ComfyUI workflow automates realistic product placement using an open-source stack. You provide up to three product shots plus a reference composition, and the graph places the products into the scene with text-aware detail and lighting that follows the reference. A Qwen3-VL stage inspects your inputs and writes a context-rich edit prompt, Flux.2 Klein 9B performs the multi-image composite, and LTX-2.3 turns the final still into a short, polished product spot. Core utility nodes like LoadImage, PreviewImage, PreviewAny, BatchImagesNode, StringConcatenate, and SaveVideo make it easy to iterate and export.

Technically, the product and reference images are loaded, batched, and auto-scaled so Flux.2 Klein receives sharp, consistent inputs regardless of your source resolution. The prompt is assembled by Qwen3-VL (guided via StringConcatenate) and passed into the Flux.2 sampler to condition placement, label fidelity, and scene lighting. After the composite is approved in PreviewImage/PreviewAny, LTX-2.3 generates a short motion segment from the final placement and SaveVideo writes the result to disk. The custom inference nodes (461b058e-342d-497e-aac1-a5fb6551e33a, fdd45ff8-8dff-43e8-b2c0-2d26053388b8, 6dbca6ef-ca4b-4af8-a99b-b834745f3a9b, 5c2c98c8-2e8f-47ca-9403-686e35a5ef94, 04e3b3b1-a711-485c-be05-f59f9507bd4e, ffacbfba-7659-45df-a3b8-7560f4800088) encapsulate the model runners and scaling/conditioning logic so you can run fully locally or on Comfy Cloud—no closed APIs.