Skip to content
Computer vision

Qwen3-VL-2B Action Genome (merged)

граф сцены по одному кадру

1 min read

Это Qwen3-VL-2B-Instruct на 2,13 млрд параметров, в который вмерджили LoRA-файнтюн под Action Genome. Модель получает один видеокадр и возвращает JSON-список объектов, с которыми связан человек, и отношений attention/spatial/contact. Вес merged-репозитория — 4,3 ГБ в bf16.

Ключевые особенности

  • На выходе только JSON: object, attention, spatial, contact по меткам Action Genome.
  • attention: looking_at, not_looking_at, unsure; contact: holding, sitting_on, carrying, touching, not_contacting.
  • LoRA r=16, alpha=32, замороженный vision tower; обучение на 3 922 кадрах из 500 видео Charades.
  • Есть отдельная adapter-only версия на 70 МБ; merged-веса — для удобства и конвертаций.
  • Не общий scene-graph: описывает только отношения человека с объектами и не работает между кадрами.

Пример

import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForImageTextToText
repo = "Ram20307/qwen3vl-2b-ag-merged"
proc = AutoProcessor.from_pretrained(repo)
model = AutoModelForImageTextToText.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto").eval()
# messages: image + PROMPT из карточки; затем model.generate(**x, max_new_tokens=400, do_sample=False)
print(out)
ПодробнееНа тесте Action Genome: JSON valid 199/199, recall 0.31, precision 0.45; spatial не оценивался.
Метрики сняты с LoRA-адаптера до merge; merged-веса отдельно не переоценивались.
Обучение: ~32 минуты на одной RTX A5000 16 ГБ, финальный train loss 0.183.
Ограничения: один кадр, узкий набор, слабые редкие классы; учебный проект, данные Charades/Action Genome — non-commercial research.
Discuss Telegram 0 22

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 4.0 GB · ready files4.9 GB
  • FP8 / INT8recommendedweights 2.0 GB2.7 GB
  • 4-bitweights 1.1 GB1.7 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 · 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.

    # Ram20307/qwen3vl-2b-ag-merged
    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 Ram20307/qwen3vl-2b-ag-merged --local-dir ~/models/qwen3vl-2b-ag-merged
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to qwen3vl-2b-ag-merged
  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 Qwen3-VL-2B-Instruct (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face
Merges 1All on Hugging Face

About the model

Ram20307/qwen3vl-2b-ag-merged
Author
Ram20307
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

220

Ask AI about this model

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

Comments

Robotics

GR00T G1

GR00T G1