Breaking

OpenVGLab Releases OmniLottie 4B to Generate Lottie Animations From Text, Images and Video

August 12, 2026 at 12:07 EDT

Media Generation
Research & Papers
Open Source

OpenVGLab has released OmniLottie, an open-source 4B model that converts text, image-plus-text and video prompts into editable Lottie JSON rather than raster video. The CVPR 2026 project combines released inference code and weights with the 2-million-animation MMLottie-2M dataset and the 900-sample MMLottieBench evaluation suite.

Mar. 3, 2026 · OpenVGLab

MULTIMODAL PROMPTS,
EDITABLE MOTION

OmniLottie generates structured Lottie JSON from text, image-plus-text or video. Its vector output preserves editable shapes, layers and motion parameters.

Input modes: text, image-plus-text and video

Released model size

Animations in MMLottie-2M

15.2 GB

Reported GPU memory use

Text-to-Lottie tokenizer ablation

Structured tokens sharply improved success

Each vertical column uses the same percentage scale.

13.4%

Raw JSON fine-tuning

97.3%

Parameterized tokens

The tokenizer replaces fragile JSON generation with compact animation commands.

Generation flow

From instruction to editable JSON

Text · Image · VideoMultimodal instruction

→

Qwen2.5-VLUnderstands the prompt

→

Lottie tokensCommands and parameters

→

Lottie JSONEditable vector animation

Why developers are interested

Open, editable and integrable

Inference code, weights and a Gradio demo
Scalable vector output instead of fixed video
Community-contributed ComfyUI integration

What still limits adoption

Reliability remains unfinished

Autoregressive decoding can produce invalid sequences
Context length constrains complex animation
Training code has not been released
The dataset is limited to noncommercial use

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.