Step-Video-T2V

What it is Step-Video-T2V is a text-to-video model from StepFun (阶跃星辰). It takes a written prompt and generates a video. It has 30 billion parameters (300亿参数). Text-to-video models generate moving images from a description, and this one was released as an open model, meaning its weights can be downloaded and run by others.

Origins StepFun released the inference code, the weights, a Turbo variant and the technical report on 17 February 2025 ^[1]. The technical report is arXiv:2502.10248 ^[2]. The model was released jointly with the carmaker Geely, together with the speech model Step-Audio ^[3].

On 17 March 2025 StepFun released Step-Video-TI2V, an image-to-video model based on Step-Video-T2V. It generates video from a starting image ^[1].

Main features The model generates videos of up to 204 frames at resolutions such as 544x992 ^[1].

Its video encoder is a deep-compression variational autoencoder (VAE). A VAE compresses data into a smaller internal representation and then reconstructs it. This one compresses video 16 times in each spatial direction and 8 times in time ^[1]. Compressing the video first makes the generation step cheaper.

The generator is a diffusion transformer with 48 layers and 48 attention heads in each layer. It uses 3D full attention, which attends across space and time together, with AdaLN-Single, QK-Norm and 3D RoPE (rotary position embeddings extended to three dimensions). It is trained with flow matching, a method that learns to move from noise to data along a path ^[1].

The model has two bilingual pre-trained text encoders, so it can take prompts in English and in Chinese ^[1].

A final training stage uses Video-DPO, which is direct preference optimisation. Direct preference optimisation trains the model on which outputs people prefer. In this case it reduces artifacts and improves smoothness ^[1].

StepFun also published Step-Video-T2V-Eval, a benchmark of 128 Chinese prompts from real users, spread across 11 categories ^[1].

Step-Video-T2V-Turbo uses inference-step distillation. Distillation trains a faster model to copy a slower one, and here it needs 10 to 15 steps, compared with 30 to 50 steps for the base model ^[1].

The model is released under the MIT license ^[1].

Place in the field Step-Video-T2V was one of the open releases that StepFun made in 2025. It came out alongside the speech models Step-Audio, in a joint release with Geely ^[3]. Its open weights, permissive license and Turbo variant make it usable by people who want to run video generation on their own hardware.