Iris-3B Turbo, OpenVINO INT8

This repo is the Iris-3B Turbo checkpoint converted to OpenVINO IR, with INT8 weights. Iris-3B is a 3B-parameter text-to-image diffusion transformer that works directly on pixels, so there's no VAE. The Turbo checkpoint is distilled to 4 steps with no CFG.

This model can be used with FastSDCPU

Inference uses only openvino, numpy, pillow and tokenizers. It doesn't need PyTorch, transformers or diffusers.

Resolution 1024×1024, batch 1 (fixed at export)
Steps 4 (distilled), no guidance scale, no negative prompt
Prompt length Up to 295 tokens (Qwen tokenizer)
Time per image 86–97 s on an Intel Core i7-12700 (CPU)
Peak RAM About 7 GB with low_memory=True, about 14 GB without
Download size 6.9 GB
Tested with OpenVINO 2026.2.0, Python 3.12, Windows 11
Downloads last month
12
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for rupeshs/iris-3b-int8-ov

Quantized
(7)
this model