HN 표시: AnimTOON – OmniLottie보다 3~4배 적은 토큰으로 SVG에 애니메이션을 적용하는 AI

3-4x fewer tokens than OmniLottie (CVPR 2026) for generating Lottie animations, running on a single consumer GPU. Model (v3): huggingface.co/srk0102200/AnimTOON-3B — now with character animation support v3 New: Multi-part SVG animation, coordinated 14-layer character idle/walk cycles, trained on Spine + DragonBones skeletal data AnimTOON is a compact, plain-text animation format designed for LLMs to generate Lottie animations with minimal tokens. Unlike existing approaches that require custom tokenizers and large GPU clusters, AnimTOON works with any LLM and runs on consumer hardware. - Character Animation: 14-layer coordinated walk/idle cycles from text - Multi-Part SVG: Animate individual parts of complex SVGs (47-part crab demo) - Spine/DragonBones Training: Model understands skeletal hierarchy (arms, legs, head, torso) - Per-Part Anchor Points: Each SVG part rotates around its own center (no more flying parts) | AnimTOON (Ours) | OmniLottie (CVPR 2026) | |---|---| | 1024 tokens / 82s / 30fps | 2001 tokens / 55s+ / 8fps | | Real SVG + per-part animation | AI-generated (not a crab) | | AnimTOON (Ours) | OmniLottie (CVPR 2026) | |---|---| | 166 tokens / 26s / 30fps | 4095 tokens / 55s+ / 8fps | | Real SVG shape + AI animation | AI-generated shape (incorrect) | Why the difference? OmniLottie tries to generate shapes AND animation in one model — leading to hallucinated shapes and token bloat. AnimTOON separates concerns: SVG provides perfect shapes, model focuses 100% on animation. | Metric | AnimTOON | OmniLottie | Raw Lottie JSON | |---|---|---|---| | Avg Output Tokens | 166-597 | 616-4095 | 18,202 | | Token Reduction | 98.8% vs JSON | ~97% vs JSON | baseline | | AnimTOON vs OmniLottie | 5-7x fewer | baseline | - | | Metric | AnimTOON | OmniLottie | |---|---|---| | Output Tokens (simple) | 166 | 616 | | Output Tokens (complex) | 597 | 4095 | | Shape + Animation Tokens | 207 (41+166) | 1113 | | Generation Time | 13-38s | 55-120s+ | | Frame Rate | 30 fps | 8 fps | | VRAM (inference) | ~5 GB | ~15.2 GB | | Model Size | 3B (LoRA) | 4B (full fine-tune) | | Training Hardware | 1x RTX 5060 Ti 16GB | Multi-GPU cluster | | Custom Tokenizer | No (plain text) | Yes (40k custom tokens) | | Accepts SVG Input | Yes | No | | Generates Shapes | No (uses SVG) | Yes (limited quality) | | Format | Plain text (any LLM) | Custom parameterized tokens | | File Size | 1-4 KB | 20-175 KB | | Format Success Rate | 100% (converter guarantees) | 88.3% | AnimTOON is a human-readable, token-efficient text format that describes Lottie animations: 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 166-token output produces a complete animated .lottie file with bounce entrance, rotation wobble, and fade-in. The same animation in raw Lottie JSON would be 18,000+ tokens. AnimTOON Pipeline +--------+ +---------+ +-----------+ +--------+ | SVG | --> | Prompt | --> | AnimTOON | --> |.lottie | | File | | Builder | | Model | | File | +--------+ +---------+ +-----------+ +--------+ | Generates only animation text (166 tokens) | v +-----------+ | Converter | | (toon_ | | animator) | +-----------+ | Combines SVG paths + model animations into .lottie file Key insight: The model only generates animation keyframes (166 tokens). Shapes come from the SVG file. The converter deterministically builds valid Lottie JSON. This separation is why we achieve 98.8% token reduction. | Approach | What Model Outputs | Tokens | Quality | |---|---|---|---| | Raw JSON | Full Lottie JSON with shapes + metadata | 18,000+ | Fragile, many format errors | | OmniLottie | Custom parameterized tokens (shapes + animation) | 486-4095 | Good but needs custom tokenizer | | AnimTOON | Plain text animation keyframes only | 166-597 | 100% valid (converter guarantees) | anim fr=30 dur=120 # framerate, duration layer Body shape # layer name + type fill #4A90D9 # color path ellipse w=0.2 h=0.2 # shape (or 'sh' for complex bezier) pos [0.5,0.5] # position (normalized 0-1) rot 0.0->0 1.0->360 ease=linear # rotation keyframes scale 0.0->[0,0] 0.2->[100,100] # scale keyframes opacity 0.0->0 0.3->100 ease=fade # opacity keyframes Supported properties: position, rotation, scale, opacity, fill, stroke, path (ellipse/rect/sh) Easing types: smooth, linear, bounce, fade This is an early research release. The model is approximately 60% through training. Full model release coming after extended training on cloud GPU. - AnimTOON format specification (complete) - Bidirectional converter: AnimTOON Lottie JSON .lottie (100% reliable) - SVG -> animated .lottie pipeline (python-lottie + model + converter) - Simple icon/logo animations: pulse, bounce, spin, fade, wobble, scale - Character animations: 14-layer coordinated walk/idle cycles - Multi-part SVG animation: 47-part crab with per-part rotation - Per-shape-group anchor point calculation (BBox centroid) - Correct color matching from text descriptions - 98.8% token reduction vs raw Lottie JSON - Trained on 100k MMLottie + 10k layer-aware + 984 Spine/DragonBones character data - No shape generation (requires SVG input) - Model output varies between runs (temperature-dependent) - Position animation on shape groups still breaks layout (rotation/scale only) - Complex multi-character scenes not yet supported - Not yet trained on facial expressions (blink/smile/talk) - v1.0: Icon/logo animation from text + SVG input - v2.0: Layer-aware animation (understands SVG structure) - v3.0 (Current): Character animation with Spine/DragonBones skeletal data - v4.0 (Next): Live2D facial expressions + anime character animation - v5.0: Full manga-to-anime pipeline (panel → animated scene) # Clone git clone https://github.com/srk0102/svg-animator.git cd svg-animator # Setup (Windows) setup.bat # Setup (Mac/Linux) chmod +x setup.sh && ./setup.sh python test_svg_pipeline.py inputs/apple.svg # Output: outputs/apple.lottie # Preview at: https://lottiefiles.com/preview 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) from src.toon_animator import animtoon_to_dotlottie_full animtoon_text = """anim fr=30 dur=90 layer Circle shape fill #FF4F59 path ellipse w=0.2 h=0.2 pos [0.5,0.5] scale 0.0->[0,0] 0.15->[120,120] 0.3->[100,100] ease=smooth opacity 0.0->0 0.1->100 ease=fade """ animtoon_to_dotlottie_full(animtoon_text, "output.lottie") # Preview at https://lottiefiles.com/preview # Download training data from MMLottie-2M (100k samples) python src/dataset_pipeline.py --limit 100000 --output data/animtoon_train.jsonl # Generate layer-aware training data python src/gen_layer_data.py # Setup Unsloth environment setup_unsloth.bat # Windows # or ./setup_unsloth.sh # Mac/Linux # Train python src/train_unsloth.py \ --data data/animtoon_train.jsonl \ --model Qwen/Qwen2.5-3B-Instruct \ --output models/animtoon-3b \ --epochs 3 \ --lora-dropout 0 python merge_lora.py # Creates models/animtoon-3b-v2-merged/ svg-animator/ src/ toon_animator.py # AnimTOON Lottie converter (core) dataset_pipeline.py # MMLottie-2M data download + conversion train_unsloth.py # LoRA training with Unsloth test_inference.py # Model inference + .lottie generation prompt_builder.py # SVG -> structured prompt gen_layer_data.py # Generate layer-aware training data svg_animate.py # SVG + AnimTOON -> animated Lottie test_svg_pipeline.py # Full SVG animation pipeline benchmark_compare.py # AnimTOON vs OmniLottie benchmark merge_lora.py # Merge LoRA into base model inputs/ # Test SVG files outputs/ # Generated .lottie files models/ # Trained model checkpoints data/ # Training data | Parameter | Value | |---|---| | Base Model | Qwen/Qwen2.5-3B-Instruct | | Method | LoRA (r=16, alpha=32) | | Training Data | 99,650 (MMLottie-2M) + 10,000 (layer-aware) + 984 (Spine/DragonBones) | | Epochs | ~2 (100k) + 3 (10k layers) + 3 (984 character) | | Hardware | 1x NVIDIA RTX 5060 Ti (16GB VRAM) | | Framework | Unsloth + Transformers | | Max Length | 1024 tokens | | Batch Size | 1 x 16 (gradient accumulation) | | Final Loss | 0.47 (100k run) / 0.24 (layer-aware run) | From 99,650 training samples (MMLottie-2M): - Average original Lottie JSON: 18,202 tokens - Average AnimTOON output: 222 tokens - Average token reduction: 98.8% - Average layers per animation: 6.2 | Method | Year | Tokens | Shapes | Animation | Custom Tokenizer | Hardware | |---|---|---|---|---|---|---| | Raw Lottie JSON | - | 18,000+ | Yes | Yes | No | Any | | OmniLottie | CVPR 2026 | 486-4095 | Yes | Yes | Yes (40k tokens) | Multi-GPU | | LLM4SVG | CVPR 2025 | ~500-2000 | Yes | No | No | Multi-GPU | | SVGDreamer | CVPR 2024 | N/A | Yes | No | N/A | GPU | | AnimTOON | 2026 | 166-597 | Via SVG | Yes | No (plain text) | 1x Consumer GPU | The AnimTOON format is designed to scale to complex character animations: # Future: anime character blink animation anim fr=24 dur=24 layer left_eye shape scale 0.0->[100,100] 0.4->[100,10] 0.5->[100,100] ease=smooth layer right_eye shape scale 0.0->[100,100] 0.4->[100,10] 0.5->[100,100] ease=smooth layer hair shape rot 0.0->-2 0.5->2 1.0->-2 ease=smooth Training on Spine/Live2D rigging data will teach the model joint constraints, coordinated multi-layer motion, and character-specific animation patterns. MIT License @misc{sivaramakrishna2026animtoon, title={AnimTOON: Token-Efficient Vector Animation Generation via Compact Text Format}, author={Siva RamaKrishna}, year={2026}, url={https://github.com/srk0102/svg-animator} } - OmniLottie - MMLottie-2M dataset and benchmark inspiration - Unsloth - Fast LoRA training - python-lottie - SVG to Lottie conversion - LottieFiles - Lottie preview and community