Skip to content
Images

Open Vision Banana 4B Seg 1-Step LoRA

сегментация по тексту за 1 шаг

2 min readupdated

Адаптер для сегментации объектов на изображении по текстовому описанию (referring segmentation): на вход подаётся картинка и фраза вроде «чёрный бильярдный шар», а на выходе получается чёрно-белая маска нужной области. Модель представляет собой LoRA (компактную надстройку над базовыми весами) от автора ApacheOne для трансформера Open Vision Banana 4B (на базе компонентов FLUX.2-klein-base-4B). В отличие от оригинальной 28-шаговой версии, этот вариант работает всего за 1 шаг деноизинга, но при этом обходит базу по точности на валидации ReasonSeg. Пригодится разработчикам компьютерного зрения и пайплайнов редактирования изображений, где критична скорость генерации масок.

Ключевые особенности

  • Генерация маски за 1 проход трансформера с обновлением Euler (от sigma 1 до 0) и CFG 1.0.
  • Ускорение этапов деноизинга и декодирования VAE примерно в 31 раз по сравнению с базовой 28-шаговой моделью (0.72 с против 22.33 с на пилотном наборе 512²).
  • На полном валидационном наборе ReasonSeg (200 изображений, 1024²) набрала 76.89 gIoU против 74.60 у оригинальной OVB 4B (+2.29 пункта) при скорости 2.46 с на изображение.
  • Обучена за 1 000 шагов на 200 изображениях из обучающей выборки ReasonSeg с параметрами rank 64 и alpha 64.

Что можно сделать

  • Автоматически выделять конкретные объекты на фото по смысловому запросу в пайплайнах на базе Diffusers и FLUX.2-klein-base-4B.
  • Использовать быструю генерацию бинарных масок как первый этап перед инпейнтингом или заменой фона в редакторах изображений.
  • Интегрировать в связку с LLM (например, Gemma 4 31B, как в бенчмарке) для превращения сложных пользовательских вопросов в точные фразы-указатели для сегментации.

Пример

pipe.load_lora_weights("ApacheOne/Open-Vision-Banana-4B-Seg-1Step-Lora", weight_name="ovb_4b_reasonseg_one_step.safetensors")
prompt = '[TASK: REFERRING_SEGMENTATION] Segment the region that best matches this description: "the black billiard ball". Paint the matching region white [255, 255, 255] and leave everything else black [0, 0, 0]. Output only the segmentation mask image.'
raw = pipe(image=image, prompt=prompt, num_inference_steps=1, guidance_scale=1.0, sigmas=[1.0]).images[0]
ПодробнееReasonSeg val (200 фото, 1024²): 76.89 gIoU, 74.79 cIoU, 2.46 с/фото (с кэшем эмбеддингов).
Пилот Open Images (10 фото, 512²): 48.96 gIoU против 46.04 у базы (0.79 с против 24.30 с).
Требования и точность: для воспроизведения бенчмарка веса LoRA переводят в float32 поверх bfloat16-пайплайна; замеры времени не включают кодирование промпта/картинки и загрузку модели.
Лицензия: FLUX Non-Commercial License (только для некоммерческого использования).
Примеры
Discuss Telegram 0 25

Will it run?

Runs in FP8 / INT8

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 34 GB · ready files39 GB
  • FP8 / INT8recommendedweights 17 GB20 GB
  • 4-bitweights 9.9 GB12 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 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 Open-Vision-Banana.

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.

    # ApacheOne/Open-Vision-Banana-4B-Seg-1Step-Lora
    # ===== 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 FunAILab/Open-Vision-Banana --local-dir ~/models/Open-Vision-Banana
    hf download ApacheOne/Open-Vision-Banana-4B-Seg-1Step-Lora --local-dir ~/models/Open-Vision-Banana-4B-Seg-1Step-Lora
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Open-Vision-Banana
  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.
  • Nunchaku (SVDQuant 4-bit)v1.2.1from January 25, 2026

    ≈3× faster and 3.5× less memory

    4-bit versions of popular image models with almost no quality loss (FP4 on RTX 50).

    How to enable: Ready weights for supported models; in ComfyUI use ComfyUI-nunchaku.

  • 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

FunAILab/Open-Vision-Banana

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

  1. FLUX.2-klein-base-9Bfamily root
  2. Open-Vision-Banana
  3. Open-Vision-Banana-4B-Seg-1Step-Lora

New posts about versions of FLUX.2-klein-base-9B (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

ApacheOne/Open-Vision-Banana-4B-Seg-1Step-Lora
Author
ApacheOne
Category
Images
Released
Posted

On NeyroLearn

Hugging Face

250

Ask AI about this model

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

Comments