Als Markdown anzeigen

Open Vision Banana

Open Vision Banana (OVB), by FunAILab, is FLUX.2 [klein] Base fine-tuned to answer perception tasks with an image. The prompt names the task and the output image encodes the answer: segmentation as flat colours, depth as grey, surface normals as RGB. The same weights keep image editing and text to image. LibreYOLO runs it through the LibreGen factory in 4B and 9B.

Aufgaben
image generation
Größen
4b, 9b at 1024 px
Installation
pip install "libreyolo[gen]"
Supportstufe
Geschwisterstufe, seit v1.8.0. Eine separate Produktoberfläche mit eigener Factory und eigenem Vertrag.
Upstream
Open Vision Banana von FunAILab, FLUX Non-Commercial License. Publikation, Quelle
Lizenzen
Code Apache-2.0, Gewichte FLUX Non-Commercial License. Kommerzielle Nutzung

Install

bash
pip install "libreyolo[gen]"

Predict

The mirrors download once on first use. LibreYOLO loads OVB with diffusers' Flux2KleinPipeline and adds no network code: what you write is the prompt and a few lines of NumPy to read the colours back.

Depth
from libreyolo import LibreGen, SAMPLE_IMAGE model = LibreGen("open-vision-banana")      # 4b; "open-vision-banana-9b" for 9bDEPTH = (    "Estimate a dense relative inverse-depth map for the input image. Produce exactly one output value "    "for every input pixel at the same pixel location, with closer points brighter and farther points "    "darker. Encode the result as neutral grayscale (R=G=B) using one consistent mapping over the "    "complete image. Preserve object silhouettes, thin structures, occlusion boundaries, smooth surfaces, "    "and small depth changes. Do not alter the scene or add texture, shading, color, text, labels, "    "legends, borders, or objects; return only the registered depth map.")result = model(DEPTH, source=SAMPLE_IMAGE, seed=0)result.generated.save("depth.png")rgb = result.generated.array                # (H, W, 3) uint8
Semantic segmentation
import jsonfrom libreyolo import LibreGen, SAMPLE_IMAGE model = LibreGen("open-vision-banana")legend = {"building": [255, 8, 232], "sky": [2, 204, 254], "water": [8, 4, 249]}prompt = (    "[TASK: SEMANTIC_SEGMENTATION] Generate a semantic segmentation visualization of the input image, "    f"using this color mapping: {json.dumps(legend, separators=(', ', ': '))} "    "Color each pixel of a listed class with its color and leave everything else black. "    "Output only the segmentation visualization image.")model(prompt, source=SAMPLE_IMAGE, seed=0).generated.save("segmentation.png")
Text to image
from libreyolo import LibreGen model = LibreGen("open-vision-banana")model("an empty tram at a cobblestone stop on a rainy night", imgsz=1024, seed=0).generated.save("tram.png")
CLI
# The CLI covers text to image only.libreyolo generate model=open-vision-banana prompt="an empty tram at a cobblestone stop on a rainy night"

source takes a path, URL, PIL.Image or BGR NumPy array. The output keeps the photo's aspect ratio, capped at one megapixel; resize it back to the photo before reading it (nearest-neighbour for masks, so colours stay exact). Without source it is a normal text-to-image model. Knobs: steps (28) and guidance (3.5), the paper's settings. offload=True keeps idle parts off the GPU.

The model was trained on many phrasings of each task. The library's Open Vision Banana guide has FunAILab's quickstart prompts and readouts for semantic, referring and instance segmentation, depth, surface normals and image editing, plus legend ensembling (several passes with shuffled colours and a per-pixel vote).

What to expect: every answer is a 28-step diffusion run, so it is slow. Use it when one model for many tasks matters more than speed. Masks have soft, slightly shifted edges after resizing; depth is relative inverse depth up to an unknown scale and shift, and normals are smooth but not metric.

Variants

SizeAliasMirrorText encoderGPU memory (library figure)
4bopen-vision-bananaLibreYOLO/LibreOpenVisionBanana4bQwen3 4Babout 16 GB
9bopen-vision-banana-9bLibreYOLO/LibreOpenVisionBanana9bQwen3 8Babout 34 GB

Both mirrors hold the paper's non-zero-shot checkpoints, trained with the benchmarks' training splits. Another OVB checkpoint saved as a diffusers pipeline (the zero-shot ones) loads with LibreGen("open-vision-banana", checkpoint_dir="path/to/folder").

Export

AufgabeONNXTorchScriptExecuTorchTensorRTOpenVINOPaddleMNNRKNNAxelerancnnTFLiteCoreMLCore AI
Image generationImage generation to ONNX: nicht unterstütztImage generation to TorchScript: nicht unterstütztImage generation to ExecuTorch: nicht unterstütztImage generation to TensorRT: nicht unterstütztImage generation to OpenVINO: nicht unterstütztImage generation to Paddle: nicht unterstütztImage generation to MNN: nicht unterstütztImage generation to RKNN: nicht unterstütztImage generation to Axelera: nicht unterstütztImage generation to ncnn: nicht unterstütztImage generation to TFLite: nicht unterstütztImage generation to CoreML: nicht unterstütztImage generation to Core AI: nicht unterstützt

Training, export and val() are not supported.

Licensing

Prüfe die Lizenz im Hugging-Face-Repository der konkreten Gewichte, die du herunterlädst. Jeder Checkpoint in der LibreYOLO-Organisation hat eine Lizenz, und sie ist innerhalb einer Familie nicht immer gleich. Dieses Repository ist die maßgebliche Quelle. Die Zusammenfassung unten beschreibt den Stand bei der letzten Verifizierung dieser Seite.

Dies ist eine Beschreibung der beteiligten Lizenzen und keine Rechtsberatung. Wenn die Antwort kommerziell relevant ist, lies die Lizenzen selbst und hole eigenen Rechtsrat ein.

Originalarbeit
Open Vision Banana, FunAILab
Upstream-Lizenz
FLUX Non-Commercial License
LibreYOLO-Code
MIT
Gewichte
FLUX Non-Commercial License, erneut unter huggingface.co/LibreYOLO veröffentlicht
Einordnung
The Open Vision Banana weights are fine-tunes of FLUX.2 [klein] Base 4B and 9B and are under Black Forest Labs' FLUX Non-Commercial License, a custom agreement that is not OSI-approved: no commercial or production use. Section 2(e) of that license asks for content filtering or human review of generated images before they are shared, and the same terms apply to the outputs. FunAILab's training code on GitHub is Apache-2.0, but LibreYOLO ports none of it: its adapter is its own MIT code that calls the Apache-2.0 diffusers Flux2KleinPipeline. LibreYOLO redistributes the paper's non-zero-shot checkpoints as complete pipelines at LibreYOLO/LibreOpenVisionBanana4b and LibreYOLO/LibreOpenVisionBanana9b, with FunAILab's transformer cast from float32 to bfloat16 (from FunAILab/Open-Vision-Banana at revision be5da66250cc385ff9c75ebbe4114e13ec305848) and Black Forest Labs' scheduler, Qwen3 text encoder, tokenizer and VAE byte-identical; the mirrors carry the license and are not gated. The zero-shot checkpoints and LoRA adapters stay on FunAILab's repository, under the same license.

Open Vision Banana is by FunAILab (OpenVisionBanana/OpenVisionBanana). The weights are under Black Forest Labs' FLUX Non-Commercial License: no commercial or production use, and section 2(e) asks for content filtering or human review of generated images before they are shared. FunAILab's training code is Apache-2.0, but LibreYOLO ports none of it.

The LibreYOLO mirrors are complete diffusers pipelines, not gated: FunAILab's transformer cast from float32 to bfloat16, the precision it runs in, and the rest of BFL's FLUX.2 [klein] Base pipeline byte-identical. Outputs match FunAILab's own quickstart bit for bit on the same seed.

Citation

@misc{jadon2026openvisionbananafully,
      title={Open Vision Banana: A Fully Open Recipe for Generative Visual Perception},
      author={Arpit Jadon and Valentinos Pariza and Jona Ruthardt and Ryousuke Yamada and A. Sophia Koepke and Andrei Bursuc and Yuki M. Asano},
      year={2026},
      eprint={2611.15844},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2611.15844},
}

Aus dem Zitierblock der Autoren unter github.com/OpenVisionBanana/OpenVisionBanana#-citation kopiert.

Mit LibreYOLO v1.8.0 verifiziert. Supporttabellen, Checkpoints und Benchmarkwerte auf dieser Seite werden aus der veröffentlichten Bibliothek und den publizierten Gewichten generiert und nicht von Hand geschrieben.