AnimTOON-3B (v3): Token-Efficient Vector Animation Generation

3-4x fewer tokens than OmniLottie (CVPR 2026) for generating Lottie animations. Now with character animation support.

	AnimTOON	OmniLottie
Tokens (simple)	166	616
Tokens (complex)	597	4095
VRAM	5GB	15.2GB
FPS	30	8
Model Size	3B LoRA	4B full
Custom Tokenizer	No	Yes (40k tokens)
Accepts SVG	Yes	No

What is AnimTOON?

AnimTOON is a compact, plain-text animation format that any LLM can generate. Instead of outputting 18,000+ tokens of raw Lottie JSON, AnimTOON describes animations in ~166-597 tokens of human-readable text.

anim fr=30 dur=120

layer Logo shape
  fill #000000
  path sh x2
  pos [0.5,0.5]
  rot 0.0->-67 0.04->46 0.14->-31 0.28->0 ease=bounce
  scale 0.0->[0,0] 0.14->[90,90] 0.28->[100,100] ease=smooth
  opacity 0.0->0 0.14->100 ease=fade

This produces a complete animated .lottie file with bounce entrance, rotation wobble, and fade-in.

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("srk0102200/AnimTOON-3B")
model = AutoModelForCausalLM.from_pretrained(
    "srk0102200/AnimTOON-3B",
    dtype=torch.float16,
    device_map="cuda"
)

prompt = "a red circle pulsing in the center with a smooth bounce"
messages = [{"role": "user", "content": f"Generate AnimTOON animation: {prompt}"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
result = tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(result)

Convert to .lottie

# Clone: git clone https://github.com/srk0102/AnimTOON.git
import sys; sys.path.insert(0, 'src')
from toon_animator import animtoon_to_dotlottie_full

animtoon_to_dotlottie_full(result, "output.lottie")
# Preview at https://lottiefiles.com/preview

Animate Any SVG

from lottie import parsers  # pip install lottie

# Convert SVG to Lottie (perfect paths)
anim = parsers.svg.parse_svg_file("your_logo.svg")
lottie_dict = anim.to_dict()

# Generate AnimTOON animations with the model
# Apply animations to the Lottie layers
# Output: .lottie file with real SVG shapes + AI animations

See full pipeline: test_svg_pipeline.py

Benchmark Results (Measured)

Same prompt, same hardware:

Test	AnimTOON Tokens	OmniLottie Tokens	Ratio
Apple logo bounce	207 (41 shape + 166 anim)	1113	5.4x fewer
Smiley face complex	597	4095	6.9x fewer
Simple ball bounce	176	616	3.5x fewer

Dataset statistics (99,650 samples):

• Average raw Lottie JSON: 18,202 tokens
• Average AnimTOON: 222 tokens
• Token reduction: 98.8%

Current Status (v3)

v3 adds character animation support trained on Spine + DragonBones skeletal data.

The model now works for:

• Icon/logo animations (pulse, bounce, spin, fade, wobble)
• Character idle/walk cycles (14 layers, coordinated)
• Multi-part SVG animation (47-part crab demo)
• Correct color matching from text descriptions
• SVG + animation pipeline with per-part anchor points

Limitations:

• No shape generation (requires SVG input)
• Model output varies between runs (temperature-dependent)
• Position animation on shape groups not yet supported
• Not yet trained on facial expressions

Training Details

Parameter	Value
Base Model	Qwen/Qwen2.5-3B-Instruct
Method	LoRA (r=16, alpha=32) merged into base
Version	v3 (final 3B Lite release)
Training Data	99,650 (MMLottie-2M) + 10,000 (layer-aware) + 984 (Spine/DragonBones)
Hardware	1x NVIDIA RTX 5060 Ti (16GB)
Framework	Unsloth
Token Reduction	98.8% vs raw Lottie JSON

Architecture: Why Animation-Only is Better

"Asking one model to draw AND animate is like asking one person to paint AND dance at the same time."

AnimTOON separates concerns:

• SVG provides shapes (perfect, no hallucination, 0 tokens)
• Model generates animation (focused, token-efficient)
• Converter merges them (deterministic, 100% valid output)

OmniLottie generates everything in one model → hallucinated shapes, token bloat (2001 tokens for a "crab" that looks like binoculars).

Links

• GitHub: github.com/srk0102/AnimTOON
• PitchHut: pitchhut.com/project/animtoon-lottie-animation
• OmniLottie (comparison): arxiv.org/abs/2603.02138
• MMLottie-2M Dataset: huggingface.co/datasets/OmniLottie/MMLottie-2M

Citation

@misc{sivaramakrishna2026animtoon,
  title={AnimTOON: Token-Efficient Vector Animation Generation via Compact Text Format},
  author={Siva RamaKrishna},
  year={2026},
  url={https://github.com/srk0102/AnimTOON}
}

License

MIT License - see LICENSE