Skip to content
Other

lmd-vitl-p8-simmim-lvd

3D-энкодер для объёмной микроскопии

2 min read

Это 3D ViT-L энкодер от ai-hhmi для объёмной электронной и expansion-микроскопии с вокселями 8 нм. Архитектура DINOv3 ViT-L с патчами 8×8×8 и 303 млн параметров: веса стартовали с 2D DINOv3 ViT-L/16 (LVD-1689M), были инфлейтнуты в 3D и дообучены SimMIM на LMD — корпусе из 243 микроскопических объёмов.

Ключевые особенности

  • 3D-энкодер для электронной и expansion-микроскопии; вход — один канал, оси (x, y, z), изотропные воксели 8 нм.
  • Архитектура DINOv3 ViT-L: 24 блока, ширина 1024, 16 голов, 3D rotary position embedding, 4 storage token; 303M параметров.
  • Инициализация из 2D DINOv3 ViT-L/16: 2D-ядро патча инфлейтнуто в 3D и уменьшено с 16 до 8 вокселей; затем 100k шагов SimMIM на LMD.
  • patch_features(x) даёт нормализованные патч-токены последнего слоя и сетку; get_intermediate_layers(...) — карты признаков из любых блоков.
  • Вход: интенсивности в [0, 1], каждая сторона кратна 8 вокселям; обучение шло на окнах 256³, другие размеры не тестировали.
  • Энкодер — обычный torch.nn.Module, можно дообучать в своём цикле; аффинные модели авторов читают блоки [7, 15, 23] через UNETR-декодер.

Пример

root = Path(snapshot_download("ai-hhmi/lmd-vitl-p8-simmim-lvd"))
config = json.loads((root / "config.json").read_text())
encoder = DinoVisionTransformer3D(**config["model"]["kwargs"])
encoder.load_state_dict(load_file(root / "model.safetensors"))
x = torch.rand(1, 1, 128, 128, 128, device=device)
tokens, grid = encoder.patch_features(x)
maps = encoder.get_intermediate_layers(x, n=[7, 15, 23], reshape=True)
print(tuple(tokens.shape), grid, [tuple(m.shape) for m in maps])
ПодробнееОтдельных бенчмарков у энкодера нет: авторы оценивали его через файнтюн для аффинностей (lmd-vitl-p8-affinity-lvd), метрики сегментации нейронов — на странице той модели и в лидерборде.
Веса: fp32, 1.21 ГБ; для запуска нужны PyTorch 2.13, NumPy, safetensors, huggingface_hub и код mia-train (коммит 0fe3872).
Ограничения: 8 нм изотропные воксели, интенсивности [0,1], стороны кратны 8; обучение на 256³, другие размеры не тестировали. Лицензия DINOv3.
Discuss Telegram 0 886

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 1.1 GB · ready files1.7 GB
  • BF16weights 0.6 GB1.1 GB
  • FP8 / INT8recommendedweights 0.3 GB0.8 GB
  • 4-bitweights 0.2 GB0.7 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 · F32 — 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.

    # ai-hhmi/lmd-vitl-p8-simmim-lvd
    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 ai-hhmi/lmd-vitl-p8-simmim-lvd --local-dir ~/models/lmd-vitl-p8-simmim-lvd
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to lmd-vitl-p8-simmim-lvd
  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

  1. dinov3-vit7b16-pretrain-lvd1689mfamily root
  2. dinov3-vitl16-pretrain-lvd1689m
  3. lmd-vitl-p8-simmim-lvd

New posts about versions of dinov3-vit7b16-pretrain-lvd1689m (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 1All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 6All on Hugging Face

About the model

ai-hhmi/lmd-vitl-p8-simmim-lvd
Author
ai-hhmi
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

8860

Ask AI about this model

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

Comments

Other

Anima

Anima