LibreOpenVisionBanana4b

Open Vision Banana 4B by FunAILab: FLUX.2 [klein] Base 4B fine-tuned to answer perception tasks with an image. The prompt picks the task; segmentation comes back as flat colours, depth as grey, surface normals as RGB. It also edits images and generates from text. This is the paper's non-zero-shot checkpoint, packaged as one complete diffusers pipeline for LibreYOLO.

from libreyolo import LibreGen, SAMPLE_IMAGE

model = LibreGen("open-vision-banana")      # downloads this repo
prompt = (
    "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."
)
model(prompt, source=SAMPLE_IMAGE, seed=0).generated.save("depth.png")

Install libreyolo[gen]. Prompts for every task and the NumPy readouts (class ids, masks, disparity, unit normals, legend ensembling): LibreYOLO's Open Vision Banana guide.

The folder also loads with diffusers alone:

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained("LibreYOLO/LibreOpenVisionBanana4b", torch_dtype=torch.bfloat16).to("cuda")

What changed from upstream

Nothing was retrained.

• transformer/: FunAILab's FunAILab/Open-Vision-Banana at revision be5da66250cc385ff9c75ebbe4114e13ec305848, folder final_models/4B/nonzeroshot/checkpoint-avg90000-92500-97500-100000_uniform/transformer, cast from float32 to bfloat16 (the precision it runs in) and otherwise unchanged. Outputs match FunAILab's quickstart, their float32 transformer loaded in bfloat16, bit for bit on the same seed.
• scheduler/, text_encoder/ (Qwen3), tokenizer/, vae/, model_index.json: byte-identical copies of black-forest-labs/FLUX.2-klein-base-4B at revision a3b4f4849157f664bdbc776fd7453c2783562f4d (Apache-2.0).

File	SHA-256
text_encoder/model-00001-of-00002.safetensors	8c0506e7f4936fa7e26183a4fd8da4e2bdbc5990ba64ae441f965d51228f36ea
text_encoder/model-00002-of-00002.safetensors	82f2bd839378541b0557bfabaf37c7d3d637071fdcb73302dedd7cf61162ce07
transformer/diffusion_pytorch_model.safetensors	4380829c290547411232937565ab7626225055ad83d77f0ea2742430f9c2b089
vae/diffusion_pytorch_model.safetensors	ca70d2202afe6415bdbcb8793ba8cd99fd159cfe6192381504d6c4d3036e0f04

Paper: Open Vision Banana: A Fully Open Recipe for Generative Visual Perception. Code: OpenVisionBanana/OpenVisionBanana.

License

The weights are derived from FLUX.2 [klein] Base and are under the FLUX Non-Commercial License (LICENSE): no commercial or production use. The bundled base components (scheduler, text encoder, tokenizer, VAE) are Apache-2.0 from Black Forest Labs (LICENSE-Apache-2.0). See NOTICE.

@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},
}