Iris1.5

Iris1.5 is a looped vision-language model built on top of:

• ByteDance/Ouro-1.4B for language generation
• NVlabs/RADIO radio_v2.5-h for image encoding

This repository does not include the full base models. It contains the trained IRIS components only: the vision-to-language projector, dual LoRA adapters, and bundle metadata needed to load them.

Examples

Tree	Food
Prompt What do you see? Answer briefly.	Prompt What food is shown in the image? Answer briefly.
Output tree	Output ramen

Quick Start

Install the requirements from this repository, then run inference with the bundled adapter weights:

python inference/infer_hf_bundle.py \
  --repo-id DavidSeyserHF/Iris1.5 \
  --bundle-subdir hf_bundle_step115000 \
  --image tree.jpg \
  --prompt "What do you see? Answer briefly."

On the first run, the base Ouro and RADIO models are downloaded and cached automatically.

License

This repository provides adapter and projector deltas only. Commercial use is restricted by the RADIO dependency license, NVIDIA Source Code License-NC: https://github.com/NVlabs/RADIO/blob/main/LICENSE

Usage is also subject to the licenses and terms of:

• ByteDance/Ouro-1.4B
• NVlabs/RADIO
• any datasets used during training

Citation

If you use this release, please cite the model repository and commit hash for reproducibility.

@misc{seyser2026iris1-5,
  title = {Iris1.5},
  author = {David Seyser},
  year = {2026},
  howpublished = {Hugging Face model repository},
  url = {https://huggingface.co/DavidSeyserHF/Iris1.5}
}