# Performance

Benchmarks on M5 Max 128GB (Apple M5 Max, 40-core GPU).

## T2V — 8-bit, 512x768

| Frames | Duration | Tokens | Denoise | Total |
|--------|----------|--------|---------|-------|
| 65 | 2.6s | 3,456 | ~80s | ~140s |
| 121 | 4.8s | 6,144 | ~80s | ~150s |
| 241 | 9.6s | 11,904 | ~178s | ~252s |

## T2V — 4-bit, 512x512

| Frames | Duration | Tokens | Denoise | VAE decode | Peak RAM | Total |
|--------|----------|--------|---------|------------|----------|-------|
| 17 | 0.6s | 768 | ~14s | ~2.4s | ~35 GB | ~50s |
| 121 | 4.8s | 4,096 | ~45s | ~8s | ~53 GB | ~101s |

## I2V — 8-bit, 768x512

| Frames | Duration | Denoise | Total |
|--------|----------|---------|-------|
| 121 | 4.8s | ~186s | ~275s |

## Pipeline Breakdown (121 frames, 4-bit, 512x512, T2V)

| Stage | Time |
|-------|------|
| Text encode (Gemma 3 12B, 4-bit pre-quantized) | ~18s |
| Load + fuse distilled LoRA + quantize transformer | ~28s |
| Load VAE decoders | ~2s |
| Denoise (8 steps, 4,096 tokens) | ~45s |
| VAE decode | ~8s |
| Audio decode + save | ~3s |
| **Total** | **~101s** |

## Notes

- Denoising time scales with token count: `frames × (height/32) × (width/32)`
- 4-bit and 8-bit have similar denoising speed — quantized matmul throughput is the bottleneck, not precision
- Peak RAM is higher than model size due to intermediate activations during denoising
- First run may be slower due to Metal shader compilation
- I2V denoising is slower than T2V at the same resolution due to the conditioning mask blending
- Pre-quantized text encoder (4-bit or 8-bit) loads significantly faster than quantizing on the fly