Skip to content
Other

Spectator: A world model for predicting brain dynamics during movie watching

3 min readupdated

Spectator is an exploratory world model from the paper State of Mind: Learning and Evaluating World Models of Naturalistic Brain Dynamics, which forecasts human brain activity while watching movies. It takes as input recent fMRI (functional magnetic resonance imaging, divided into 1000 Schaefer-1000 regions) recordings and video stream features from the TRIBE v2 model, forming a latent state of 2048 numbers and rolling out predictions forward. Under the hood, it relies on a linear encoder trained using the JEPA principle (latent-space prediction without raw data reconstruction) and frozen, along with a transition matrix calculated in closed analytical form. Unlike approaches like TRIBE v2 that rely solely on stimuli, Spectator accounts for the current state of the viewer and noticeably outperforms competitors on the BrainStateBench benchmark. The release will be useful for neuroscientists and computational neuroscience researchers working with the Algonauts 2025 and HCP datasets.

Key features

  • Hybrid forecasting: combines the viewer's fMRI history (7.5, 120, and 180-second options) with movie context features from TRIBE v2.
  • Analytical solution: the transition matrix to the next state is calculated in closed form, and a linear probe is used for decoding into 1000 brain regions.
  • Ready checkpoints and metrics: weights for the Algonauts 2025 and HCP datasets (184 participants, 7T fMRI) are published in the repository, along with saved BrainStateBench results.
  • CC BY-NC 4.0 license: the model and used TRIBE v2 features are available for non-commercial research.

What you can do

  • Reproduce experiments: run a 40-step fMRI rollout and recalculate linear probes on the Algonauts 2025 (Friends S6 and Movie10 films) or HCP datasets.
  • Compare your architectures: upload custom models to the ready-made benchmark-results storage and get paired 95% confidence intervals against Spectator.
  • Fine-tune the transition matrix: use the frozen encoder (encoder.pt) as a starting point for analytical fitting on new recordings.

Example

folder = "outputs/spectator-v1/algonauts/model"
cfg = load_config(f"{folder}/config.yaml")
model, _ = load_world_model(cfg, f"{folder}/best.pt", "cpu")
ref = torch.load(f"{folder}/reference-rollout.pt", weights_only=True)
with torch.no_grad():
    states = model.rollout(ref["brain"], ref["stimuli"], horizon_trs=40)
Read moreDetails
On the Algonauts 2025 test (180 s history): Forecast — 16.0%, State — 5.5%, Updating — 20.3%, Pearson correlation r = 0.349 (versus 11.6%, 2.6%, 12.7% and r = 0.294 for Linear forecaster).
On HCP (184 viewers): Forecast 6.6% on new content and 9.6% on new viewers, State 6.2%, Updating 15.7% and 18.2%.
File sizes: the main Algonauts checkpoint weighs 181 MB, the HCP version is 253 MB, and the benchmark storage is 503 MB.
Limitations: the checkpoints do not contain optimizer state (gradient fine-tuning cannot be continued), and the fMRI recordings themselves are not included in the repository.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.1 GB · ready files0.6 GB
  • FP8 / INT8recommendedweights 0.1 GB0.6 GB
  • 4-bitweights 0.0 GB0.5 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.

    # stanjsx/Spectator
    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 stanjsx/Spectator --local-dir ~/models/Spectator
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Spectator
  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 Spectator (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

stanjsx/Spectator
Author
stanjsx
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

Other

EpiClock

EpiClock