1. *_FP16

What it is: Full version of the model in 16-bit floating point precision.

Pros: Highest image fidelity and feature completeness for this version. Widely supported by all GPUs.

Cons: Uses more VRAM than BF16 or DMD2, slightly slower.

Use case: When quality matters most and you have enough VRAM (8GB+ recommended). Great for detailed, high-res renders.

2. *_BF16

What it is: Same model, but in bfloat16 precision.

Pros: Similar image quality to FP16, but may perform better on GPUs/TPUs optimized for BF16 (e.g., newer NVIDIA RTX 30/40 series, some cloud accelerators).

Cons: Not as universally supported as FP16; benefits only on certain hardware.

Use case: Best if your hardware has native BF16 acceleration—slightly faster and can be more numerically stable than FP16 in training/inference.

3. *_DMD2

What it is: Distilled with Distribution Matching Distillation (v2) for speed and lower VRAM usage.

Pros: Lighter, faster, needs less VRAM (can run on 6GB cards or lower). Still maintains most of the quality of the full model.

Cons: Slightly less precise in fine details or extremely complex scenes.

Use case: Ideal for lower VRAM setups, faster batch generation, or when speed is more important than absolute maximum detail.

If you want a simple rule of thumb:

Maximum quality & compatibility: *_FP16

Quality + hardware-optimized speed: *_BF16

Speed & low VRAM: *_DMD2