Open-Weights Image-to-Video Models: Licences, Output, and What Is Downloadable

You have a still image on your hard drive and want three seconds of motion from it. You do not want to upload the picture to a company server, pay per clip, or hand over copyright. What you need is an open-weights image-to-video model with a downloadable licence that lets you run the generation on your own computer. Three models are downloadable right now: Stable Video Diffusion, LTX-Video, and Wan 2.1. For each one, know the licence, the commercial-use terms, the VRAM you will need, and what the output actually looks like in practice versus a hosted service.

Before you pick a model, understand the licence gap. Open-weights means the model file is downloadable from Hugging Face but the training data, the code used to train it, and the data-processing pipeline are not provided. Open-source offers the full reproducibility chain. Every model covered here is open-weights. None of them lets you reproduce the training run yourself.

That distinction matters for commercial use because the licence is the only legal document governing what you can do with the downloaded file. If a model were truly open-source, the licence might still restrict use, but at least you could fork the training code. With open-weights you get the model or you do not. There is no middle ground.

What You Get

Stable Video Diffusion from Stability AI is the lowest-barrier entry to local image-to-video. The model card, dated 2023-11-21, describes two versions: SVD at 14 frames and SVD-XT at 25 frames, both at a native resolution of 576×1024. The default frame rate in the reference code is 7 fps, so a 14-frame clip runs about 2 seconds and a 25-frame clip about 3.6 seconds. You configure the amount of motion with a motion bucket ID from 1 to 255. There are no camera controls, no region masks, and no start-and-end-frame support. Audio generation is absent.

Licence and Training Data

The licence is the Stability AI Non-Commercial Research Community Licence. Commercial use, selling the output, using it in client work, embedding it in a product, is not permitted. Check the licence URL on the Hugging Face model card for the exact terms. Training data disclosure states WebVid-10M and a curated internal dataset, though the full composition is not itemised.

Hardware Reality

Documented VRAM requirement in the model card reflects ideal batch-size-1 inference. Real usage in ComfyUI with the VAE decode and an upscaling node pushes peak VRAM 2-4 GB higher. On an 8 GB card, run the base SVD pipeline. On a 12 GB card, you have room for SVD-XT plus a basic upscaler. Expect the clip to degrade after about 4 seconds of motion: subject drift, limb mutation, background sliding. That is the trade-off: you get something running on modest hardware, but the motion coherence drops off before the clip ends.

What You Get

LTX-Video from Lightricks is the model to reach for when you need speed and a commercial-friendly licence. The model card, dated 2024-11-22, states a native clip length of 121 frames at 768×512 or 97 frames at 512×768. At 24 fps default, that is roughly 5 seconds and 4 seconds respectively. Inference time is listed as fast on consumer hardware; the model card emphasises speed as a design goal. There are no documented motion controls, no camera controls, no start-and-end-frame support, and no audio generation.

Licence and Training Data

The licence is the Lightricks Community Licence. Commercial use is permitted for individuals and organisations with less than $1 million annual revenue. Above that threshold you need an enterprise licence. If you are a solo filmmaker or small studio, this is the most permissive commercial option among the three models covered here. Training data is disclosed as public datasets including Vimeo-90K, Panda-70M, and HD-VILA-100M.

Hardware Reality

Documented VRAM requirement assumes batch-size-1 inference. In practice, a full ComfyUI pipeline with upscaling and VAE decode can exceed the stated figure by 2-4 GB. On a 12 GB card, run the model at 768×512 without upscaling. On a 16 GB card, you have room for upscaling to 1080p. The failure mode to watch for is motion collapse: the model biases toward low-motion outputs because they produce fewer visible artefacts during training. If you push the motion bucket ID high, expect subject drift rather than physically consistent movement.

What You Get

Wan 2.1 from Alibaba Tongyi Lab offers the highest native resolution among the open-weights image-to-video models in this comparison. The model card, dated 2025-02-25, lists two parameter sizes: the I2V-14B model at 720P (1280×720) and the I2V-14B-480P model at 480P (832×480). Both produce 81 frames at 16 fps default, which is about 5 seconds of video. There are no motion controls, no camera controls, no start-and-end-frame support, and no audio generation.

Licence and Training Data

The licence is Apache 2.0. Commercial use is permitted. Training data is not itemised in the model card. The 14B parameter count means this model is large.

Hardware Reality

The documented VRAM requirement for the 720P model is high. On a 24 GB card, run inference at 720P without upscaling. Real usage with the full ComfyUI pipeline and a VAE decode will push peak VRAM 2-4 GB above the documented minimum. The 480P model fits on a 16 GB card more comfortably. The failure mode to watch for is aspect-ratio distortion: the model is trained on specific aspect ratios and produces artefacts when you feed a source image cropped to an unexpected shape.

Pick Wan 2.1 when source-image fidelity at clip end matters most. The 720P native resolution means you do not need to upscale from a low base. The trade-off is VRAM cost and inference time. On a consumer 24 GB card you can run it, but you will not be able to run anything else at the same time.

The two models people most often compare are LTX-Video and Wan 2.1. The distinction is straightforward. LTX-Video is optimised for speed on consumer hardware and uses a single-size architecture. Wan 2.1 offers multiple parameter-size options, 1.3B and 14B, with the 14B model delivering higher native resolution at the cost of VRAM. The Lightricks Community Licence permits commercial use under a revenue cap. Apache 2.0 permits commercial use without a cap. If you need 720P native resolution and have 24 GB of VRAM, pick Wan 2.1. If you need fast iteration on a 12 GB card and the Lightricks Community Licence terms work for you, pick LTX-Video.

Stable Video Diffusion and LTX-Video are distinguished by licence. SVD uses the Stability AI Community Licence, which is non-commercial. LTX-Video uses the Lightricks Community Licence, which permits commercial use under a revenue cap. SVD is known for low VRAM entry and ComfyUI integration. LTX-Video is known for fast inference on consumer hardware. If you need commercial use, LTX-Video is the only option of the two. If you are experimenting and have an 8 GB card, start with SVD.

The image-to-video open-weights vs open-source distinction is the single most important legal checkpoint before you download anything. An open-weights model gives you the weights file. That is it. You cannot inspect the training data, you cannot reproduce the training run, and you cannot verify the data-processing pipeline. If the licence says non-commercial, you are bound by it. If the licence says Apache 2.0, you have commercial freedom but no insight into what the training data contains.

The risk is not legal action from the model provider. The risk is platform rejection. A model trained on data you cannot audit may produce output that a platform's detection system flags as synthetic, and you have no way to explain why the output is there because you do not know what the training data looks like.

The single most common failure when running any of these models locally is VRAM exhaustion crash during the VAE decode stage, not during the main inference. The model's latent-space compression hides the peak memory usage until the decode step. You see the inference complete, then the process dies with an out-of-memory error. Reduce the batch size, reduce the number of frames, or add a memory optimisation such as sequential offloading in ComfyUI.

The second most common failure is motion collapse: the model generates a nearly static clip despite a motion bucket ID set to 255. This happens because the model biases toward low-motion outputs. Restart the generation with the same settings. The noise seed matters. If you get a static clip, change the seed and retry.