Skip to content
Other

Sightline-heads

Draft heads for rapid SVG and Lottie generation acceleration

2 min readupdated

The repository contains a set of trained weights (draft heads) implementing the Sightline method, which accelerates vector graphics generation models without losing output quality. The method relies on speculative decoding (where a small auxiliary model quickly proposes a token window and the base model verifies them in a single step). The repository features configurations for model families such as OmniSVG, InternSVG, SVGen, HiVG, OmniLottie, and StarVector with block sizes of 16 and 32. This tool is useful for researchers and engineers working on optimization and vector generation acceleration.

Key features

  • Generation acceleration: Draft heads significantly reduce the generation time of SVG and Lottie animations.
  • Wide support: Ready-to-use weights for 11 different setups and architectures (OmniSVG, SVGen, StarVector, and others).
  • Safetensors format: All weights are saved in the modern format precisely matching the paper and benchmark results.
  • Flexible verification windows: Options are available with a block size of 32 (for the Sightline method) and a block size of 16 (for the baseline DFlash and DDTree approaches).

What you can do

  • Accelerate SVG generation: Plug the heads into corresponding target models (such as OmniSVG-8B or StarVector) for fast text-to-vector and image-to-vector conversion.
  • Generate Lottie animation: Use the heads for the OmniLottie-3B model within vector animation generation pipelines.
  • Run benchmarks: Compare performance against DFlash and DDTree methods using scripts from the official Sightline repository.

Example

Downloading the required weights via the command line and running evaluation for text models:

hf download lizaiyuan0804/Sightline-heads --include "omnisvg-8b/text-to-svg-b32/*" --local-dir heads
python3 eval/eval_text.py --head heads/omnisvg-8b/text-to-svg-b32 --designs D1-top16-N127 
    --sets bench --modes greedy,sample --time-ar --out $SIGHTLINE_OUT/ours
Read moreDetails
Architecture: 5 transformer layers utilizing hidden states of the base model without its own embedding.
Training: on MMSVG-2M and MMLottie-2M datasets under the CC BY-NC-SA 4.0 license.
Weight size: from 0.7 GB to 3.2 GB depending on the target model (in bfloat16 format).
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 3.0 GB · ready files3.8 GB
  • FP8 / INT8recommendedweights 1.5 GB2.2 GB
  • 4-bitweights 0.8 GB1.4 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # lizaiyuan0804/Sightline-heads
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download lizaiyuan0804/Sightline-heads --local-dir ~/models/Sightline-heads
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Sightline-heads
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Sightline-heads (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

lizaiyuan0804/Sightline-heads
Author
lizaiyuan0804
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments