AI product animation
Thomas Lundström claims that a single reference image of a product is enough to produce professional-looking, customizable product ads in Kling 3.0. The workflow generates a clean start frame and end frame, then lets Kling animate between them. He says this keeps the product and its label consistent and leaves the model less room to hallucinate.
After a week of hands-on testing, the author says he found a workflow that keeps the product consistent in Kling 3.0 product animations. This follows his earlier first-impressions video, where he promised to test text labels and warping. The only input needed is a simple reference image of the product.
The core move is to create two still frames, a start frame and an end frame, and have Kling 3.0 animate between them. The author says this gives a clean starting point and end point where the product is consistent, so the model has less room to hallucinate during the animation.
The pipeline runs from a plain product shot to a finished clip. The AG1 example: a white-background product image, then a Pinterest style reference, then his custom GPT image prompter to write the image prompt, then Nano Banana Pro in Higgsfield for the start frame. The end frame comes from the start frame plus the product image, with a prompt for a similar environment. A regular ChatGPT session, given both frames and told it is for Kling 3.0, writes the motion prompt. Kling 3.0 in Higgsfield's video tab produces the clip.
The author treats generation as iterative. He says you may need to generate several times and tweak the prompt, for example adding an explicit instruction that the green powder must not spill around the bag after early generations showed weird spilling.
He reports the results with their flaws. The AG1 clip keeps the product label completely intact at the end. Small text warping occurs mid-animation, but he says motion blur can disguise it. A small zoom and speed adjustment in editing software can smooth it further.
He chose the Ordinary and Liquid Death animations as deliberate stress tests, because the small text and the complicated front graphic should be very hard for AI. The purpose was to see whether the labels would fall apart or stay intact with this workflow.
A perfect loop comes from reusing the same two frames in two separate clips with start and end swapped. The first clip goes wide shot to close-up, the second goes close-up to wide shot, and each has a different animation in between. Played back to back in an editor, they loop seamlessly.
His conclusion is that this is an easy way to get professional-looking custom product animations without ruining the label text. He says it is not perfect, but you can make impressive results quickly from just a reference image. He also promotes a Skool community that hosts all his prompts and resources.
Because the end frame is generated from the original product image, every clip lands on a clean, correct label. Whatever warping happens is confined to the middle of the clip, where movement masks it.
The start frame can leave elements out and the end frame can add them back. In the AG1 example, the start frame excludes the scoop and bottle, and the end frame includes them. The motion prompt then describes them sliding in from the right, so the frame difference drives the choreography.
The author delegates motion prompting to a general chatbot. He uploads both frames to regular ChatGPT, says he is using Kling 3.0, and describes the desired animation, instead of writing the Kling prompt by hand.
Fixing an artifact means adding a negative constraint to the motion prompt, such as no spilling of the powder around the bag. It does not mean changing the frames.
For the candy loop, he built the frames in reverse order. He first generated the close-up end frame with the eyes closed. He then referenced that image and the original product image, and asked for a zoomed-out version with the original product in place of the one with eyes, which became the wide-shot start frame.
He deliberately picked label-heavy products (the Ordinary's small text, Liquid Death's complex front graphic) to test the workflow where AI is weakest, instead of showcasing easy products.
Light post-processing is part of the recipe. A small zoom and a speed adjustment in an editor are suggested to make the generated clip look smoother.
«So, it gives less room for the model to hallucinate during the animation.»
— 01:09
«I chose these two products specifically because they should be really difficult for AI to deal with with the small text»
— 04:58
«Now, that's a quite solid animation in my opinion, keeping the product label completely intact at the end.»
— 04:21
«there were some small warping of the text. But as the product is moving, that can also be disguised in the motion blur itself.»
— 04:28
«To make an animation that loops perfectly, we can use the two same starting and end frames for two separate animations.»
— 05:22
«Is it perfect? No. But you can create some pretty cool stuff quite quickly using this workflow and just a reference image of your product.»
— 06:49
Reception
Mostly warm and engaged, with viewers thanking the creator, joining the community, and asking for more tutorials, though a few pointed critiques about the animations' conversion value, their purpose, and the Skool community's pricing temper the praise.
A concise, honest walkthrough of a start-frame/end-frame workflow for label-safe product animations in Kling 3.0, including a loop variant. Its value is the frame-anchoring logic and the reported prompt tweaks. The full prompts sit behind the author's community.

07:23