PixArt-LCM XL-2 1024-MS (0.6 B params, OpenRAIL++-M)

SDXL-class 1024-px images in four diffusion steps.

• 1024-px in one shot. Pure-Transformer latent model spits out 1024 × 1024 frames straight from the prompt—no upscaler or second pass needed.

• Lightning inference. Just 4 steps → ≈0.5 s on an A100 and 3.3 s on a free-tier T4, 4-8 × faster than vanilla SDXL.

• Tiny footprint. Only 0.6 B parameters; FP16 needs ~6 GB and 8-bit fits < 8 GB VRAM—single 8 GB GPUs or beefy laptops can serve full-res images.

• Drop-in pipeline. PixArtAlphaPipeline.from_pretrained("PixArt-alpha/PixArt-LCM-XL-2-1024-MS", torch_dtype=torch.float16)—supports Diffusers, ONNX/OpenVINO, CPU offload, and torch.compile speed-ups.

• License quickie. CreativeML OpenRAIL++-M: free commercial use but standard “no porn/hate” content rules.

Why pick it for Norman AI?

Near-SDXL quality at sub-second latency on a single A10G lets us power real-time hero art, on-device previews, or an edge “speed tier” in our generative stack—no H100 bill required.

response = await norman.invoke(
    {
        "model_name": "stable-diffusion-v1-5",
        "inputs": [
            {
                "display_title": "Prompt",
                "data": "A cat playing with a ball on mars"
            }
        ]
    }
)