Skip to content
Video

TriCosmos-2.5-predict

multi-sensor road forecasting in adverse weather

3 min readupdated

A world model from SORA-Lab that predicts the next 2 seconds of vehicle motion across three sensors simultaneously: a standard RGB camera, a thermal imager, and a LiDAR (laser radar) depth map. It takes 5 observed frames at 10 Hz as input and outputs 20 future frames in fog, rain, snow, or at night. Built on the NVIDIA Cosmos-Predict2.5-2B base and fine-tuned via LoRA (low-rank adapters) on the DSERT-RoLL dataset, its core trick is packing all three streams into a single video frame arranged as a 2×2 grid. Useful for autonomous driving researchers and developers running simulations and planning in harsh weather.

Key features

  • Unified 2×2 grid: a 1280×704 frame is split into four 640×352 tiles — RGB at top-left, thermal at top-right, LiDAR depth at bottom-left, and an empty black square at bottom-right.
  • Higher LiDAR accuracy: joint training across all three modalities cuts depth prediction mean absolute error (MAE) down to 4.68 m — 2.1 m better than a comparable LiDAR-only LoRA model (6.81 m).
  • Built for harsh weather: trained on 35 sequences from DSERT-RoLL covering dense fog (23 clips), heavy snowfall, rain, and night scenes.
  • Ready-to-use merged checkpoint: the repository includes tricosmos_ema_bf16.pt (3.9 GB, bf16 format) along with the experiment.json configs needed to assemble the model with LoRA parameters.

What you can do

  • Simulate adverse-weather driving scenes: plug it into cosmos-predict2.5 pipelines to generate 2-second multi-sensor scenarios in fog, snowfall, and nighttime conditions.
  • Benchmark multimodal world models: compare joint RGB, thermal, and depth predictions against single-sensor baselines on the DSERT-RoLL validation split.
  • Fine-tune on custom sensor logs: use the published train_tri_lora.yaml recipe and 2×2 tile-packing layout to adapt Cosmos-Predict2.5-2B to your own automotive sensor recordings.

Example

import torch
sd = torch.load("tricosmos_ema_bf16.pt", map_location="cpu", weights_only=False)
sd = sd.get("model", sd) if isinstance(sd, dict) else sd
lora = [k for k in sd if "lora" in k.lower()]
print(len(sd), "tensors,", len(lora), "LoRA tensors")
Read moreDetails
Metrics (DSERT-RoLL validation, 2 s horizon): RGB PSNR — 21.71, Thermal PSNR — 24.61, LiDAR MAE — 4.68 m (vs. 5.71 m for simply repeating the last frame and 6.81 m for LiDAR-only LoRA).
Training: LoRA rank 16, alpha 32 on attention and MLP layers; 500 steps in bf16 on 93-frame clips; generation evaluated at 35 sampling steps and guidance 7.
Limitations: each sensor runs at half-tile resolution (640×352), simple last-frame repetition for LiDAR remains more accurate at close ranges under 10 m, and the authors plan to release the full inference code after paper peer review.
License: NVIDIA Open Model License (based on Cosmos-Predict2.5-2B).
Discuss Telegram 3 47

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 15 GB · ready files21 GB
  • FP8 / INT8recommendedweights 9.6 GB14 GB
  • 4-bitweights 7.1 GB11 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 5080 · BF16 — from RTX 4090

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

This is a LoRA: counted together with its base model Cosmos-Predict2.5-2B.

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.

    # SORA-Lab/TriCosmos-2.5-predict
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    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 nvidia/Cosmos-Predict2.5-2B --local-dir ~/models/Cosmos-Predict2.5-2B
    hf download SORA-Lab/TriCosmos-2.5-predict --local-dir ~/models/TriCosmos-2.5-predict
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Cosmos-Predict2.5-2B
  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.

  • SageAttention 3v2.2.0from October 28, 2025

    even faster on RTX 50 / Blackwell

    FP4 attention for Blackwell. Faster than SageAttention 2, but the authors suggest v2 when accuracy matters.

    How to enable: Build from the sageattention3_blackwell folder of the repo; in ComfyUI via KJNodes.

    Earlier generations (2)
    • SageAttention 2++ — Quantised (INT8/FP8) attention with no visible quality loss. The main speed-up for image and video generation.
    • SageAttention 1 — First version: INT8 for Q·K. Still works, but 2++ is faster and more accurate.
  • 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 (3)
    • 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.
    • xFormers — The older way to speed up attention, works almost everywhere. Sage or Flash are faster where available.
  • cache-dit (DBCache)v1.5.6from November 20, 2026new release

    1.5–3× at nearly the same quality

    Reuses block outputs between diffusion steps. Supports most DiT models.

    How to enable: pip install cache-dit; one line cache_dit.enable_cache(pipe) for diffusers.

    Earlier generations (2)
    • MagCache — Skips steps by magnitude of change. More accurate than TeaCache, no calibration.
    • TeaCache — The first popular diffusion step cache. Works, newer methods are more accurate.
  • LightX2V step distill0.6.1from November 9, 2026new release

    5–10× fewer steps

    Distilled LoRAs and an engine: video in 4–8 steps instead of 30–50.

    How to enable: Load their step-distill LoRA, CFG 1.0, 4–8 steps.

  • 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

Requires the base model

nvidia/Cosmos-Predict2.5-2B

This model is a LoRA / adapter. Download the base model first, then load this one on top.

New posts about versions of Cosmos-Predict2.5-2B (quants, finetunes, LoRAs) will show up under the bell.

Fine-tunes 12All on Hugging Face

About the model

SORA-Lab/TriCosmos-2.5-predict
Author
SORA-Lab
Category
Video
Released
Posted

On NeyroLearn

Hugging Face

473

Ask AI about this model

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

Comments