Product videos from photos with local AI: image-to-video for shops and ads on your own GPU

Product clips from the photos you already have

An online shop can turn its existing product photos into short video clips on its own workstation with open image-to-video models. Wan 2.2 animates a packshot into a clip of about five seconds at 480p or 720p under the Apache 2.0 licence, and LTX-2.5 adds sound and multi-shot sequences under a licence that is free below USD 10 million of annual revenue. One NVIDIA RTX PRO 6000 Blackwell Workstation Edition or Max-Q with 96 GB meets the 80 GB the publisher states for the large Wan model, and a 24 GB RTX PRO 4000 Blackwell meets the 24 GB stated for the small one.

This guide is for shop owners, e-commerce and marketing managers and agencies that produce clips for product pages, marketplaces and social ads, and for the IT person who will run the machine. It covers the models and their licences, the steps from photo to clip, how a whole catalogue runs as a queue, and which hardware fits which volume.

Some terms first. Image-to-video, often shortened to I2V, means the model takes a still image as the first frame and invents the motion that follows: a slow turn of the product, a camera move, light passing over a surface. A packshot is the clean product photo you already use in the shop. The prompt is a sentence that tells the model what should move. An upscaler raises the finished clip to a higher resolution.

Our guide to the e-commerce AI server covers product texts, search, chat and image editing, and our guide to AI product photos covers cut-outs and packshots; this article adds the video step.

Open image-to-video models for product clips

Hugging Face model cards of Wan-AI and Lightricks and the LICENSE file of LTX-2, read 30 October 2026; Wan's frame count and rate from its configuration files as quoted in our Wan 2.2 server article.

Wan 2.2, released by the Wan team on 28 July 2025, comes in two sizes for this job. The A14B image-to-video model is a mixture of two experts with about 14 billion parameters each, one for the layout and one for the details, and its card states that its single-GPU 720p command "can run on a GPU with at least 80GB VRAM". The card also says the size setting fixes the area of the clip, "with the aspect ratio following that of the original input image", and that the model can generate a video "solely from the input image", without a prompt. TI2V-5B, the small model, works from text or an image at 720p and 24 frames per second; its card says it runs on "a GPU with at least 24GB VRAM" and "can generate a 5-second 720P video in under 9 minutes on a single consumer-grade GPU", without naming the card.

LTX-2.5 from Lightricks has a 22-billion-parameter transformer and a custom Gemma 4 12B text encoder, and generates synchronised video and sound. Images are passed with a frame index, frame 0 being the first frame, and native multi-shot generation keeps characters, setting and style across cuts. Its 4K example produces 3840 × 2176 with two spatial upscaling passes. The card gives no memory figure, only low-memory tips: converting the 16-bit transformer to FP8 when it loads, and moving parts to the processor.

Licences for shop and client work

Product clips are commercial use, so the licence is the first filter. Wan's cards state: "The models in this repository are licensed under the Apache 2.0 License", which permits commercial use with no territory or revenue condition, and the Wan team claims no rights over the generated content.

The LTX-2.x Community License, dated 11 August 2026, defines Commercial Entities as "Entities with annual revenues of at least $10,000,000" and requires them "to obtain a paid license for any use", except non-commercial purposes such as testing and evaluation in a non-production setting. For the threshold, "all subsidiaries, affiliates, and companies under common Control shall be considered collectively". Lightricks "claims no rights in the Output", and its acceptable use policy prohibits placing generated content in any context "without expressly and intelligibly disclaiming that the information and/or content is machine generated". A shop below the threshold may use LTX-2.5 in production at no cost under that licence; a larger company needs the paid agreement.

Two models are out: HunyuanVideo 1.5 and MiniMax H3, whose licences exclude the EU; we do not install them.

The AI Act's deep-fake rule also covers objects such as products, and Article 50 applies from 2 August 2026. The Act defines a deep fake as AI-generated or manipulated content that "resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful", and Article 50(4) requires deployers to disclose such content. Whether a given product clip falls under it, and how to label it, is a question for your legal counsel; our article on the EU AI Act gives general information.

From packshot to finished clip in ComfyUI

ComfyUI ships ready workflows, called templates, for each Wan 2.2 job, among them "Wan 2.2 14B Image to Video", "Wan 2.2 14B First-Last Frame to Video" and "Wan 2.2 5B Video Generation", and Lightricks states that official LTX-2.5 templates ship in ComfyUI as well. ComfyUI's App Mode shows such a workflow as a simple form, so a marketer picks the photo, writes one sentence and starts the job without seeing the node graph.

1. Choose the photo. A packshot with the whole product in frame and a calm background gives the model the least to invent. Its shape sets the format: a horizontal photo for 16:9, a vertical one for 9:16 social clips.
2. Describe the motion. "Slow 360-degree turn on a white table" or "camera moves in, soft light passes over the lid". Wan also runs without a prompt.
3. Fix the end if needed. The first-last-frame template takes a second photo as the last frame, for a box that opens or a product that ends on its label.
4. Check every clip. The model redraws every frame. None of the model cards we read promises that printed text, logos or the shape of the product stay unchanged, so compare each clip with the photo before it goes online. A clip that changes the product shows the customer something you do not sell.
5. Finish. Upscale with SeedVR2 (Apache 2.0, native in ComfyUI) to full HD, then add text, music and the logo in the editor. Our guide to AI video upscaling compares the options.

Short clips work best for this. For the A14B models Wan's default is 81 frames at 16 frames per second, about five seconds, which matches the length of a product loop or a social ad scene; longer spots are cut together from several clips.

You can order the workstation alone, the setup of ComfyUI, the Wan 2.2 templates and an upscaler alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

Queues and batch runs for a whole catalogue

A shop with hundreds of products runs the clips as a queue. ComfyUI describes "asynchronous queueing" and a local API for integrating workflows into applications, so the shop system or a script can submit one job per product with the photo and a prompt. Each card works on one clip at a time; two cards finish two clips at once.

How many clips an hour that gives is not published for any RTX PRO card. Our Wan 2.2 server article estimates an upper bound from NVIDIA's peak compute rate for the A14B model at Wan's default 40 steps, computed for the text-to-video model, whose two experts are the same size as the image-to-video ones: at least 17.2 minutes per 720p clip and at least 4.4 minutes per 480p clip on one RTX PRO 6000 Server Edition, so at most 3 and 13 clips per hour per card. Real runs are slower. By our estimate, 500 products at 480p therefore take at least 36.7 hours on one card and at least 9.2 hours on a server with four cards. Fewer steps, the 5B model or compressed weights shorten that, by amounts no publisher has measured on these cards, so time a few of your own clips on the delivered system before you plan a catalogue.

Offloading, which Wan's 80 GB figure assumes, moves weights into the computer's main memory. By our arithmetic in the same article, each A14B job in 16-bit keeps about 69 GB there, so a workstation that runs it on two cards needs main memory for both jobs on top of the operating system. Our workstation configurations follow NVIDIA's rule of at least twice the GPU memory: 192 GB next to one 96 GB card and 384 GB next to two.

Which workstation or server for which catalogue

Memory decides which model and which precision fit on the card; the card's compute decides how many clips an hour it delivers. The memory figures below come from the model cards and from the file sizes listed on Hugging Face; working memory for the frames comes on top.

memory statements from the Wan and LTX-2.5 model cards and ComfyUI's Wan 2.2 tutorial; file sizes from Hugging Face (Comfy-Org Wan 2.2 repackaged, Lightricks/LTX-2.5), read 30 October 2026; card memory from NVIDIA's datasheets.

Entry. Wan states at least 24 GB for TI2V-5B with offloading, and ComfyUI's tutorial says the 5B model "should fit well on 8GB vram with the ComfyUI native offloading". The RTX PRO 4000 Blackwell (24 GB, 145 W, single slot) meets Wan's figure; the RTX PRO 4500 Blackwell (32 GB, 200 W) leaves more room for 720p frames. Both suit a shop that wants to try clips on its best-selling products first.

Mid. The RTX PRO 5000 Blackwell with 72 GB is below Wan's 80 GB for the 16-bit A14B command. ComfyUI's repackaged FP8 files are smaller: the two 14B experts at 14.3 GB each, the FP8 text encoder at 6.7 GB and the VAE at 0.25 GB come to about 35.6 GB by our arithmetic. For LTX-2.5, the NVFP4 transformer of 18.7 GB, which Lightricks lists with Blackwell GPUs, and the INT8 text encoder of 15.4 GB come to about 34 GB. By our usable-memory rule (90 per cent of the memory the driver reports, less 3 GiB) the 72 GB card offers about 61.5 GiB, so it holds either set with room for the frames.

96 GB. One RTX PRO 6000 Blackwell, as Workstation Edition (600 W) or Max-Q (300 W), meets Wan's 80 GB as stated. In 16-bit the A14B weights total 64.4 GiB by the arithmetic of our Wan server article, and LTX-2.5's 16-bit transformer (42 GB) and text encoder (26.3 GB) come to about 68 GB. A second card lets two clips run at once. Two Workstation Edition cards are the realistic maximum for that edition; three or four Max-Q belong only in a chassis built for them with strong directed airflow.

Large. A server with four RTX PRO 6000 Server Edition cards runs four clips at once in an overnight queue and draws about 3.4 kW by our estimate, which needs a 32 A or three-phase feed in a server room. An L40S server (48 GB per card) holds TI2V-5B and the FP8 A14B files by the same file arithmetic; LTX-2.5's NVFP4 file is listed for Blackwell GPUs, and the L40S is a generation older. Our video generation GPU requirements guide sizes team servers further, and our AI video workstation for marketing compares the workstation tiers.

The software, and where your product data stays

When we set up the software, the workstation gets ComfyUI, under GPL-3.0 and actively developed as of October 2026, the Wan 2.2 models and templates and SeedVR2 for upscaling. LTX-2.5 and Wan2GP, a one-window interface, are not part of that setup; each has its own licence, under which a company that meets the terms can install it itself.

Product photos, prompts, unreleased products and finished clips stay on the workstation. The internet is needed only to download models and updates; generating a clip runs offline. Store the prompt, the seed and the model version with every clip, so that a clip can be regenerated after a product change.

We supply workstations from one RTX PRO 4000 to two RTX PRO 6000 Workstation Edition or Max-Q and servers for catalogue runs; send us the number of products, the formats and how often the catalogue changes, and we propose a configuration and the models for it.

FAQ