koala-lightning-700m

KOALA-Lightning 700M (≈ 700 M params, license unspecified)

SDXL-quality images at SSD-speed.

• Big images, blink-fast. Spits out a 1024 × 1024 frame in ≈ 0 .66 s with 10 denoise steps on an RTX 4090—about 4 × quicker than plain SDXL.

• Runs on 8 GB cards. 3 GB checkpoint and compressed 700 M-param U-Net mean even a 3060 Ti can do full-res inference without OOM.

• Distilled from SDXL. Keeps the look by siphoning self-attention features from SDXL and training on high-res LAION-POP captions.

• License caveat. No explicit license listed—assume research-only until the authors clarify.

Why pick it for Norman AI?

KOALA-Lightning gives us near-SDXL quality at a fraction of the latency and VRAM, perfect for a “speed” tier in the Norman generative stack—think instant hero art, real-time preview, or edge deployments without H100 price tags.

response = await norman.invoke(
    {
        "model_name": "koala-lightning-700m",
        "inputs": [
            {
                "display_title": "Prompt",
                "data": "A cat playing with a ball on mars"
            }
        ]
    }
)