What is Step-Video-TI2V?
Step-Video-TI2V is an open-weight image-to-video generation model from StepFun. It accepts a reference image together with a text prompt, then generates a video that uses the image as the starting visual condition while following the prompt’s instructions for movement, action, or scene development.
The model is part of StepFun’s Step-Video family and is based on Step-Video-T2V. The “TI2V” designation reflects its text-and-image-to-video workflow: the image supplies visual context, while text describes what should happen. For example, a user could provide an image of a landscape and prompt the model to add slowly moving clouds, or submit a character illustration and ask for a particular motion or camera-like change.
StepFun’s official repository describes the model as having 30 billion parameters. The official Hugging Face model card lists the checkpoint at approximately 29 billion parameters in BF16 format, so the exact figure depends on how the model is described and measured. The released weights and inference code are available under the MIT license.
How the model uses images, prompts, and motion control
The reference image is incorporated as the first-frame condition. This means the generated sequence is intended to develop from the supplied visual rather than being created solely from a text description. Text then provides direction about the desired action, motion, or transformation.
Step-Video-TI2V also exposes a motion-score parameter. This control is designed to adjust the amount or intensity of motion in the generated result. A lower setting may be useful when preserving the composition and appearance of the reference image is more important, while a higher setting can be appropriate when the goal is more visible movement. The supplied research confirms the existence and purpose of this control, but does not establish a universal mapping between particular score values and specific motion outcomes.
In practical terms, the model is best understood as a controllable animation and visual-transition system rather than a general-purpose text or coding model. It is designed to synthesize video frames from visual and textual conditions, not to answer questions, produce structured business data, or execute external tools.
Verified specifications
| Specification | Verified information |
|---|---|
| Provider | StepFun |
| Model family | Step-Video |
| Primary task | Text-guided image-to-video generation |
| Parameters | 30B-class according to StepFun; approximately 29B on the Hugging Face model card |
| Inputs | Reference image and text prompt |
| Maximum documented output | Up to 102 frames |
| Motion control | Motion-score parameter |
| Weights and code | Publicly released for self-hosted inference |
| License | MIT |
| Release date | March 17, 2025 |
The 102-frame figure is the maximum documented configuration in the official repository, technical report, and model-card materials reviewed. It should not be interpreted as a guaranteed duration in seconds because the actual playback length depends on the frame rate and generation settings used by a particular workflow.
Access, hardware, and deployment requirements
Step-Video-TI2V is primarily a self-hosted model. StepFun provides an official repository, while the model weights are available through Hugging Face. An official ComfyUI integration is also available for users who prefer a visual workflow. The supplied research does not verify a general-purpose hosted API with published per-generation pricing for this model.
The hardware requirements are a significant part of the practical evaluation. Official documentation describes approximately 75 GB of peak GPU memory on a single GPU for documented 102-frame configurations. A four-GPU configuration is described as requiring approximately 64 GB per GPU. These figures indicate that the model is aimed at well-equipped workstations, servers, or research infrastructure rather than ordinary consumer laptops.
Memory use can vary with resolution, frame count, precision, optimization, and implementation details. The official figures should therefore be treated as documented reference requirements, not as a guarantee that every setup will use exactly the same amount of memory.
Main strengths
Reference-image control
The model’s most important distinction is its use of a reference image as the first-frame condition. This gives users a direct way to start from an existing composition, character, product image, or scene instead of relying only on a text prompt. That makes the model relevant to image animation, visual prototyping, and experiments that require a particular starting appearance.
Open-weight flexibility
Because the weights and inference code are publicly released under the MIT license, users can inspect, run, and integrate the model in their own environments, subject to their responsibility for deployment and use. Self-hosting can be valuable for researchers who need control over their infrastructure or want to experiment without sending source images to a third-party hosted service.
An explicit motion control
The motion-score parameter provides a documented control for adjusting generated movement. This is useful when a workflow needs to explore different levels of animation while retaining the same reference image and prompt. It offers a more direct experimental handle than simply rewriting the prompt and hoping the motion changes as intended.
Limitations and trade-offs
Large resource footprint
A roughly 29B-to-30B parameter video model with documented memory requirements measured in tens of gigabytes is not a lightweight option. Setup, GPU availability, memory management, and inference time can all be meaningful constraints. The supplied evaluation information rates its speed at 2 out of 5 and cost at 3 out of 5; these are editorial scores, not StepFun-published benchmarks or prices. They reflect the practical trade-off between open access and the infrastructure needed to run the model.
No verified model-specific hosted pricing
No official model-specific input price, output price, per-video price, or hosted subscription price was verified in the supplied research. As a result, there is no reliable recurring or per-generation price to quote. The financial cost of using the model depends primarily on the hardware and hosting arrangement selected by the user.
Not a general-purpose assistant
Step-Video-TI2V does not provide text output, structured output, tool use, web search, or conventional coding capabilities. It is not intended to replace a language model for writing, analysis, programming, or workflow automation. Its output is video, and its inputs are text plus an image.
Undocumented context and API limits
A conventional language-model context window and maximum output-token limit do not apply to this model in the usual sense. No model-specific context length or maximum output-token limit was published in the reviewed sources. The relevant documented output constraint is the maximum of 102 frames for the referenced configurations. Other limits, such as supported resolutions or practical generation durations under different settings, should be confirmed from the implementation being used rather than assumed.
Supported modalities and capabilities
Step-Video-TI2V supports text input and image input. Its direct output is video. The supplied research does not verify native audio input, video input, audio output, music generation, speech generation, or video editing of an existing video as separate capabilities.
Its reasoning and coding capabilities are not applicable in the way they are for language models. The model is not documented as a reasoning model, programming assistant, or tool-calling system. It should be evaluated by the quality and controllability of its generated motion, visual consistency, and adherence to the reference image and prompt rather than by text benchmarks.
Best use cases
- Animating a still illustration, concept image, or designed scene.
- Creating short visual prototypes from a reference frame and written motion direction.
- Research into open-weight video generation and image-conditioned synthesis.
- Testing how motion intensity changes through the motion-score control.
- Building self-hosted or ComfyUI-based image-to-video workflows.
- Experimenting with video generation where an MIT-licensed model and local processing are useful requirements.
The model is especially suitable when the user values control over the starting image and is prepared to manage the necessary compute. It can also be useful when local or self-hosted processing is preferable to sending reference material to a hosted provider.
When to choose Step-Video-TI2V
Choose Step-Video-TI2V when you need an open-weight image-to-video model, have access to substantial GPU memory, and want to experiment with reference-image animation and explicit motion control. Its MIT license, public weights, and official inference resources make it more adaptable than a closed hosted endpoint for users who can handle deployment themselves.
A managed video-generation service may be more appropriate when you need a simple web interface, predictable per-generation pricing, low operational overhead, or infrastructure that can scale without maintaining large GPUs. A smaller or optimized video model may also be preferable when speed and memory efficiency are more important than using this particular open-weight checkpoint.
Step-Video-TI2V is not the right choice for text generation, coding, tool-using agents, audio-video editing, or workflows that require a documented commercial API price. It is also a poor fit for low-memory environments unless the selected implementation provides additional optimizations that meet the user’s requirements.
Bottom line
Step-Video-TI2V is a specialized StepFun model for turning a reference image and text prompt into a short video. Its defining advantages are open weights, an MIT license, image-conditioned generation, documented output of up to 102 frames, and a motion-score control. Its defining costs are operational: the model is large, the documented hardware requirements are substantial, and no model-specific hosted pricing was verified.
For researchers, developers, and creators with suitable GPU infrastructure, it offers a flexible foundation for self-hosted image-to-video experimentation. For users who mainly want fast, simple, and predictably priced video creation, a managed or smaller alternative may be more practical.