Free and Local Image-to-Video Generation: Models, Tools, and Hardware

You have a source image. You want it to move. And you do not want to pay for a cloud subscription or hand over your content to a platform that owns the output terms. The question is whether free and local image-to-video generation can deliver usable clips on the hardware you already own, and if not, which alternative costs you nothing. Every capability, limit, licence term, and availability claim in this guide carries its date in the sentence that states it and names the documentation it came from.

Image-to-video generation takes a single still image and produces a short moving clip, typically 2 to 10 seconds. The models do not understand the scene. They predict pixel motion from patterns learned during training. The thing newcomers most often get wrong is assuming the output will be physically consistent. Objects morph. Limbs drift. Backgrounds slide independently of the foreground because there is no persistent 3D world model, only a statistical guess at what motion looks like. That constraint is structural and will not change in the next generation of models. Every tool below operates inside it.

• Core Architecture: Latent diffusion with temporal attention layers. Source image is encoded into latent space, noise is injected, and frames are denoised across time. No persistent 3D world model exists.
• Output Standard: Generated clips are almost universally MP4 container, H.264 or H.265 encoding, at 24/25/30 fps. FFmpeg is the universal post-processing tool for frame extraction, concatenation, and format conversion.
• Local Interface Standard: ComfyUI node graph is the standard local-generation front-end for open-weights video models. Hugging Face is the distribution mechanism for model weights.
• Legal Boundary: Open-weights models carry either a non-commercial or commercial licence. The distinction is embedded in the model's terms and affects resale and broadcast use. Fully AI-generated video lacks copyright registration under US Copyright Office guidance (88 FR 16190).
• Transparency Obligation: EU AI Act Article 50(2), effective 2026-08-02, requires deployers of synthetic video to disclose the output is artificially generated. Platforms including YouTube, TikTok, and Meta require labelling of AI-generated content.

You fall into one of five groups. Pick the one that describes what you need to do, not the one that sounds most interesting.

The tinkerer who wants to run models on their own GPU goes to the open-weights models page. That page answers which open-weights image-to-video models accept image input, which licence each carries, and what the documented VRAM requirement is at 720p. It separates Stable Video Diffusion 1.1 from LTX-Video 0.9.5 by licence: SVD is non-commercial under the Stability AI Non-Commercial Research Community Licence, while LTX-Video is Apache 2.0. It distinguishes Wan 2.1 by its multiple parameter-size options against LTX-Video's single-size architecture optimised for speed. It tells you that documented VRAM minimums reflect ideal batch-size-1 inference and that real usage with the full ComfyUI pipeline, including VAE decode and upscaling, can exceed the stated figure by 2 to 4 GB.

The person who wants zero cost and no hardware goes to the free tiers page. That page answers which services offer a genuine free tier, what the limits are, and where the catch sits. It names the difference between the marketing term "unlimited" and the fair-use policy buried in terms of service.

The person who wants the local workflow but lacks the GPU goes to the cloud GPU rental page. That page answers which rental services support the required drivers, how the inference time per second of video compares to local generation, and what the real cost is when queue wait time on free tiers is factored in.

The person who wants the step-by-step ComfyUI setup goes to that page. It answers which nodes to install, what order to connect them, and which failure modes appear at each stage: VRAM exhaustion crash, prompt ignored, and content-moderation false positives.

The person evaluating commercial use for broadcast or advertising goes to the legal and compliance page. That page answers the exact disclosure obligations under the EU AI Act, the copyright registration status of AI-generated video under US Copyright Office guidance, and the terms-of-service restrictions on commercial use for each model.

Running an open-weights image-to-video model locally means you command the input, the output, and the timeline. No queue priority. No content moderation false positive from a platform's safety filter. No throttling. You also own the hardware cost and the failure modes.

Wan 2.1 (Alibaba Group)

Wan 2.1 is released under an Apache 2.0 licence, which permits commercial use. It accepts JPEG and PNG source images. Native clip length is 5 seconds. Native resolution tops out at 720p (1280×720). Motion controls are limited to a text prompt: no start or end frame, no motion strength slider, no camera control, no region masks. The model is available in multiple parameter sizes (1.3B and 14B), which gives you a choice between speed and output quality on a consumer GPU. The weights are distributed on ModelScope and can also be accessed via the Alibaba Cloud API. The free demo on ModelScope offers a way to test output before committing to a local download. Documentation as of 2026-09-26 does not specify a maximum input resolution.

LTX-Video 0.9.5 (Lightricks Ltd.)

LTX-Video 0.9.5 is released under the Lightricks Community Licence, which is a non-commercial licence. Commercial use requires a separately purchased licence. The model accepts JPEG and PNG source images at a documented training resolution of 768×512 px. Native clip length is configurable from 2 to 6 seconds. Motion controls are text-prompt-only. There are no start or end frame inputs, no motion strength slider, no camera control, and no region masks. The model is distributed on Hugging Face and has a ComfyUI node available. The architecture is a single-size design optimised for fast inference on consumer hardware. If you are running this on a consumer GPU, expect the VAE decode stage, not the main inference, to be the VRAM peak that causes crashes.

Stable Video Diffusion 1.1 (Stability AI Ltd.)

Stable Video Diffusion 1.1 is released under the Stability AI Non-Commercial Research Community Licence. Commercial use is not permitted. The model accepts JPEG and PNG source images at a documented resolution of 1024×576 px. Native clip length is 2 to 4 seconds, producing 14 or 25 frames at a configurable 3 to 30 fps. Motion controls consist of a motion strength parameter (motion_bucket_id, 0 to 255) and a frame-rate setting. There are no start or end frame inputs, no camera control, and no region masks. The weights are on Hugging Face and a ComfyUI node is available. The entry point for VRAM is lower than Wan 2.1 at 14B, but the trade-off is shorter clip length and weaker camera-motion control granularity than proprietary models like Runway Gen-4 or Kling 2.1.

ComfyUI is the node-graph interface that serves as the standard local-generation front-end for open-weights image-to-video models. When a new open-weights model ships, ComfyUI workflows are its primary local deployment method. The pipeline structure is stable: source image to pre-processing (resize, crop, aspect-ratio adjustment) to model inference to post-processing (frame interpolation, upscaling, format encoding).

The person who needs a step-by-step ComfyUI setup should go to that dedicated guide. It answers which nodes are required for each model, what order they connect in, and which failure modes appear at each stage. The VRAM exhaustion crash is the most common failure during local generation. It almost always occurs during the VAE decode stage, not during the main inference. The documented VRAM requirement for a model like LTX-Video 0.9.5 is lower than the real peak usage when the full pipeline, including upscaling and frame interpolation, is active. That guide explains how to measure your actual peak and where to put the memory-saving nodes.

The distinction between ComfyUI and Automatic1111 matters. ComfyUI has native video-model support. Automatic1111 is image-generation-focused and requires extensions for video. If you are starting from zero, ComfyUI is the correct interface for image-to-video work.

Consumer GPU hardware is the constraint that determines whether a model runs at all. Go to the hardware requirements page to learn which GPU can run which model. That page answers the specific VRAM requirements at 720p for each open-weights model, and it distinguishes between documented minimums and real peak usage. Documented minimums reflect ideal batch-size-1 inference. Real usage with the full ComfyUI pipeline including VAE decode and upscaling can exceed the stated figure by 2 to 4 GB.

The page names which consumer cards can run Wan 2.1 at 14B, which can run it at 1.3B, and where the line is drawn for LTX-Video 0.9.5 and Stable Video Diffusion 1.1. It covers the trade-off between motion magnitude and source fidelity: stronger motion increases the risk of subject drift and artefacting, a structural consequence of predicting pixel displacement from a single reference frame. It does not quote live retail prices for GPUs because those change too fast, but it describes the price band each card sits in.

If you want the local workflow but lack the GPU, cloud GPU rental bridges the gap. Go to the cloud GPU rental page. That page answers which rental services support the required drivers for ComfyUI nodes, what the inference time per second of video looks like on rented hardware, and what the real cost is when queue wait time on free tiers is factored in.

The claimed inference speed on a rented GPU is not the same as the total turnaround time. API-documented inference time excludes server-queue wait time, which can multiply total turnaround by 2x to 10x during peak usage. The page names the services that offer a free tier and what the queue priority looks like on each. It also covers the content moderation strictness on the rental platform, which may reject source images that a local model would accept.

If you want zero cost and no hardware, the free tiers of proprietary services are the practical path. Go to the free tiers page. That page answers which services offer a genuine free tier, what the daily or monthly generation cap is, and where the catch sits. The catch is not always visible in the marketing. Unlimited generations on subscription tiers carry throttling, reduced priority, or lower resolution for high-volume users. The fair-use policy is buried in the terms of service.

The page names the free tier availability for each service as documented on 2026-09-26. Runway Gen-4 offers limited free generations on its Basic plan. Kling 2.1 offers daily free credits in its web app. Luma Dream Machine offers limited free generations per month. Pika 2.2 offers limited free generations per day. Hailuo AI offers limited free generations per day. Sora has no free tier. Seedance 1.0 offers free generations in the Jimeng app but requires a Chinese phone number. Veo 3 from Google DeepMind offers limited free generations in VideoFX in select countries. These figures are revised annually; the provider publishes the current allocation.

The proprietary models that offer a free tier each have a different constraint set. Understanding the difference between them determines whether the free tier is usable for your specific source image and output requirement.

Runway Gen-4 (Runway AI, Inc.)

Runway Gen-4 accepts JPEG, PNG, and WebP source images. Native clip length is 5 or 10 seconds. Native resolution is 720p or 1080p. Audio generation is not supported. Motion controls include a motion strength slider (0 to 10), camera control sliders for pan, tilt, zoom, and roll, and a start frame image input. There is no end frame input and no region masks. The free tier is available on the Basic plan with limited generations. Commercial use is permitted on paid plans; free tier outputs have usage restrictions per Runway terms. Documentation as of 2026-09-26 does not specify a maximum input resolution.

Kling 2.1 (Kuaishou Technology)

Kling 2.1 accepts JPEG and PNG source images. Native clip length is 5 or 10 seconds. Native resolution is 1080p. Audio generation is not supported. Motion controls include a motion strength parameter (0 to 1), a start frame image input, an optional end frame image input, and camera control presets for pan, tilt, zoom, and roll. There are no region masks. The free tier offers daily free credits in the Kling web app. Commercial use is permitted subject to Kuaishou terms. Documentation as of 2026-09-26 does not specify a maximum input resolution.

Luma Dream Machine (Luma AI)

Luma Dream Machine accepts JPEG and PNG source images. Native clip length is 5 seconds. Native resolution is 720p or 1080p. Audio generation is not supported. Motion controls include a start frame image input, an optional end frame image input, and camera control presets for pan, tilt, zoom, and orbit. There is no motion strength slider and no region masks. The free tier offers limited generations per month. Commercial use is permitted subject to Luma terms. Documentation as of 2026-09-26 does not specify a maximum input resolution.

Pika 2.2 (Pika Labs, Inc.)

Pika 2.2 accepts JPEG and PNG source images. Native clip length is 4 seconds. Native resolution is 720p or 1080p. Audio generation is supported, either generated with the video or from an uploaded audio file for lip-sync. Motion controls include a motion strength slider (0 to 10), camera control presets for pan, tilt, zoom, and roll, and region masks for selective motion within the frame. There is no start or end frame input. The free tier offers limited generations per day. Commercial use is permitted subject to Pika terms. Documentation as of 2026-09-26 does not specify a maximum input resolution.

Hailuo AI (MiniMax)

Hailuo AI accepts JPEG and PNG source images. Native clip length is 6 seconds. Native resolution is 720p or 1080p. Audio generation is not supported. Motion controls are text-prompt-only. There are no start or end frame inputs, no motion strength slider, no camera control, and no region masks. The free tier offers limited generations per day. Commercial use is permitted subject to MiniMax terms. Documentation as of 2026-09-26 does not specify a maximum input resolution.

The specification table on this subject is full of numbers that look precise but are not comparable.

Claimed Maximum Clip Length Versus Usable Clip Length

Many models advertise 10-second generation. Motion coherence degrades visibly after 4 to 6 seconds. The gap is caused by the model's attention mechanism losing track of the source image over long sequences. The models that support start-and-end-frame conditioning, Kling 2.1 and Luma Dream Machine, produce more consistent longer clips because the second frame anchors the endpoint.

Claimed Resolution Versus Effective Resolution

A model may output 1080p files, but independent frame analysis shows effective detail closer to 720p upscaled. The gap is roughly one resolution tier, caused by the latent-space compression used in the generation pipeline. LTX-Video 0.9.5 trains at 768×512, which is below 720p. Stable Video Diffusion 1.1 trains at 1024×576. Wan 2.1 tops out at 720p. If you need true 1080p output, you are looking at the proprietary models. Runway Gen-4, Kling 2.1, Luma Dream Machine, Pika 2.2, and Hailuo AI all claim 1080p native resolution. Test the effective detail yourself before relying on it for client delivery.

Claimed Inference Speed Versus Real-World Queue Time

API-documented inference time excludes server-queue wait time, which can multiply total turnaround by 2x to 10x during peak usage. On a free tier, your queue priority is lowest. The inference time per second of video is the number that matters for interactive work. The total turnaround time is the number that matters for production deadlines.

Documented VRAM Requirement Versus Actual Peak VRAM Usage

This is the gap that crashes local generation. Documented minimums reflect ideal batch-size-1 inference. Real usage with the full ComfyUI pipeline including VAE decode and upscaling can exceed the stated figure by 2 to 4 GB. The hardware requirements page gives the specific numbers for each GPU and model combination. Those figures shift with driver updates, model quantisations, and ComfyUI node optimisations. Measure your actual peak before buying hardware.

The licence on an open-weights model determines whether you can sell the output, use it in client work, or incorporate it into a product. The distinction is embedded in the model's terms and will not change within the useful life of this content. Wan 2.1 is Apache 2.0, which permits commercial use. LTX-Video 0.9.5 is the Lightricks Community Licence, which is non-commercial; a separate commercial licence is available from Lightricks. Stable Video Diffusion 1.1 is the Stability AI Non-Commercial Research Community Licence, which permits non-commercial research only.

Proprietary models each have their own terms. Runway Gen-4 permits commercial use on paid plans; free tier outputs have usage restrictions per Runway terms. Kling 2.1 permits commercial use subject to Kuaishou terms. Luma Dream Machine, Pika 2.2, Hailuo AI, and Seedance 1.0 all permit commercial use subject to their respective terms of service. Sora permits commercial use subject to OpenAI terms. Veo 3 permits commercial use via the paid API tier; free tier outputs are subject to Google's terms.

Watermarking ranges from visible to invisible metadata to none. The choice affects platform compliance. A visible watermark affects aesthetics but satisfies platform labelling requirements where a metadata-only watermark may not. The free tiers page names the watermarking policy for each service as documented on 2026-09-26.

If you plan to animate a real person, get consent. The legal and compliance page covers the exact obligations, but the rule is simple: you need documented permission from the person in the source image before you generate motion from their likeness. No model licence waives that requirement.

Every image-to-video model has documented failure modes. Knowing them before you pick a tool saves you from blaming the wrong thing.

Subject drift, the person or object in the source image gradually transforming into something else, is caused by the model's temporal attention losing the conditioning signal from the source image. It is worst on long clips with high motion strength. Kling 2.1 and Luma Dream Machine mitigate it with end-frame conditioning. Neither eliminates it.

Limb mutation, arms, legs, or fingers warping, multiplying, or detaching during motion, is caused by the model's lack of a skeletal or physical model of the subject. No current model solves this. Region masks in Pika 2.2 can limit the area where mutation occurs, but they do not prevent it inside the masked region. Background sliding, the background moving independently of the foreground subject, is caused by the model failing to infer a consistent 3D scene structure from the single image. It is worse on models with no camera control, where the only motion parameter is a text prompt. Identity shift, a face changing ethnicity, age, or gender during the clip, is caused by the model interpolating toward the distribution of faces in its training data rather than anchoring to the source.

Flicker, rapid frame-to-frame changes in brightness, colour, or texture, is caused by inconsistent latent-space decoding across frames. It is more common on models with fast inference. Motion collapse, the model generating a nearly static clip despite a prompt requesting significant movement, is caused by the model's bias toward low-motion outputs, which have fewer visible artefacts during training.

VRAM exhaustion crash is the failure mode that stops local generation mid-inference. It is almost always the VAE decode stage, not the main inference, that causes the peak. The hardware requirements page explains how to measure your peak and where to put memory-saving nodes. Content-moderation false positive, a benign source image rejected by the safety filter, is caused by overly broad classifiers that flag skin tones, certain object shapes, or patterns that resemble prohibited content. On a free tier, you cannot circumvent this. Running locally, the ComfyUI setup page explains which nodes to remove or modify.

Start with your hardware. If you own a consumer GPU with at least 8 GB of VRAM, begin at the open-weights models page. It tells you which models accept image input, what licence each carries, and what the documented VRAM requirement is at 720p. From there, the hardware requirements page tells you whether your specific card can run the model you want. If your card cannot run it, the cloud GPU rental page tells you which service to use and what the real cost is after queue wait time.

If you own no GPU and want to spend nothing, begin at the free tiers page. It names which services offer a genuine free tier, what the daily or monthly generation cap is, and which services flag generated content for platform labelling. If you are producing for social-media short-form video and need speed, Pika 2.2 with its region masks and audio generation may be the right choice. If you need camera control, Runway Gen-4 or Kling 2.1 offer the most granular options on their free tiers.

If you are evaluating for commercial use, begin at the legal and compliance page. It names the exact disclosure obligations under the EU AI Act Article 50(2), the copyright registration status under US Copyright Office guidance, and the terms-of-service restrictions on commercial use for each model. Do not assume a free tier permits commercial use. Do not assume an open-weights model permits commercial use. Read the licence.