AI compositing
By chaining his two earlier ComfyUI workflows, Ricky says you can do AI compositing on video: replace an object (here a sedan with an SUV) and get matching relighting, shadows and moving reflections. First he builds one AI-composited reference frame. Then he uses that frame to drive a video pass.
The video claims that combining the author's previous two workflows (Qwen image compositing and the cleanup/AI-compositing video workflow) goes well beyond object swaps. The same setup can relight, add shadows and add moving reflections, and the author calls all of it 'super simple and straightforward'.
The problem is a replacement object that doesn't fit the plate. The SUV's lighting, perspective and time of day differ from the footage. The claimed result is reflections that match the environment, a day-to-night relight, and the car's lights turned on.
Stage one is a single-frame reference. Build a clean plate of the background, cut out the subject, make a deliberately rough composite, and let the Qwen image model fix it. The video says the model copes even when the composite is slightly off, and that it warps the car, wheels included, to match the scene's perspective and adds lights and reflections.
Stage two is the video pass, anchored on that reference image. The reference image is described as a 5-second video clip to make the prompt. The original car is segmented and its mask is altered. The result is described as working out of the box, with moving shadows and reflections.
The result is presented as a strong first attempt rather than a finished one. Wheel motion is the acknowledged gap, to be handled by running the inpaint workflow again with a specified wheel speed or a request that the wheels move.
The rough composite only has to communicate placement and scale. The model is credited with working out perspective, lighting and reflections itself, so it needs no careful matte or colour matching.
The clean plate comes from a single frame fed to Flux with the instruction to remove the car. That frame becomes the background for the reference composite.
The video mask is deliberately distorted with blockify mask and grow mask so it no longer looks like a car. The stated reason is that the Wan video model tries to add a car if the alpha has the shape of the car.
The video prompt is written from the reference image, not from the source footage. A vision-language model is asked to describe the reference as a 5-second clip of the SUV driving at night on the streets of Tokyo.
The reference frame is the creative anchor. Lighting, time of day and the car's lit-up state are decided in the single-frame stage, and the video stage propagates them with moving shadows and reflections.
Little in the video is new tooling. The contribution is the chaining of two existing workflows: a Qwen image composite as the reference, then the cleanup workflow reused for video compositing.
«was playing around and using my previous two workflows I realized we can do a lot.»
— 00:14
«And even if it's slightly off, it's super smart to figure things out.»
— 02:33
«So, as you can see, it uh warped the car a bit to match the perspective.»
— 04:00
«describe the image as if it's a 5-second long video of the SUV driving at night on the streets of Tokyo.»
— 05:11
«using the blockify mask and grow mask it will enlarge and uh change the shape of the car so it doesn't look like a car»
— 06:01
«So when we cue this, you should get something straight out of the box working just like your reference image.»
— 06:26
«for the wheels, you can do an inpaint again of this workflow»
— 06:43
Reception
Viewers are strongly appreciative of the channel's compositing-focused tutorials and the creator's helpful replies, with only mild critiques about the missing workflow and technical limits like 8-bit output and troubleshooting issues.
A compact demonstration of a two-stage pipeline: a Qwen-built reference frame feeds a Wan-based video pass for object replacement with relighting and moving reflections. It builds on the author's earlier workflows and ends with the first-attempt result, leaving wheel motion for a further inpaint pass.

07:08