Fully-sovereign text-to-video on Apple Silicon (Wan2.1), parity to diffusers

Contribution

Claimed. We demonstrate a fully-sovereign text-to-video inference pipeline for Wan2.1-T2V-1.3B on Apple Silicon in which a prompt string is processed end-to-end in-process — tokenizer -> UMT5 text encode -> UniPC multistep schedule -> DiT denoise with classifier-free guidance (CFG) -> 3D-VAE decode -> frames — with every linear running through Apple's Accelerate BLAS and no third-party ML framework (no diffusers, no torch) in the path. Against an fp32 diffusers reference with matched seed and schedule, this pipeline reaches per-step latent cosine 1.0 across the denoise schedule and output frames cosine 1.0 at PSNR 73.3 dB.

NOT claimed. This is a numerical-parity / correctness-validation result, not a novel algorithm and not a performance result. We do not claim throughput or latency competitiveness — the 3D-VAE causal convolution is a naive implementation and dominates wall time. We do not claim perceptual generation quality beyond the measured numeric parity (cosine/PSNR against the reference is not a human-quality judgment). We do not claim coverage of larger Wan variants, longer clips, or other video models/backends. The altitude here is an engineering re-implementation validated for bit-faithful agreement with an established reference, not a research advance in modeling.

Methods

Results

Threats to validity / Limitations

• Sample size is a correctness smoke test, not a benchmark. The parity is established over a single prompt / bounded decode and a single denoise schedule. This validates numerical agreement, not behavior across a prompt distribution, seeds, resolutions, or clip lengths.
• Single run, no variance. Cosine/PSNR figures are reported as point values with no repetition statistics, confidence intervals, or seed sweep recorded in source notes. There are no wall-time numbers recorded at all, so even the qualitative "VAE conv is slowest" claim is not backed by tabulated timings in the source.
• Parity != quality. Cosine 1.0 / PSNR 73.3 dB measures agreement with the fp32 diffusers reference. It says nothing about perceptual or generation quality; if the reference itself is imperfect, this pipeline faithfully reproduces those imperfections.
• Reference-bounded. All claims are relative to one specific fp32 diffusers configuration with matched seed/schedule. The exact reference version and the exact scheduler/step settings are not recorded in source notes — so the parity is only as trustworthy as that under-specified oracle.
• Scope / quantization bounds. Validated only for Wan2.1-T2V-1.3B at fp32 on Apple Silicon. No claim holds for larger Wan variants, quantized paths, longer clips, or non-Accelerate backends.
• Hardware/version under-specification. The exact Apple Silicon chip, RAM, OS, toolchain, and reference framework versions are not recorded in source notes, which limits independent reproducibility.
• Evidence level. The source describes the parity result but does not, in these notes, document an independent second-party reproduction with full version pins. A skeptical reader should treat the level as "validated against an under-specified reference" until the pins and a re-run are recorded.

Reproducibility / Provenance

• Build / run commands: standard Swift release build and test.
• Sovereignty: zero third-party ML framework calls (Accelerate BLAS only; no diffusers, no torch).
• Version pins: recent Swift toolchain and recent diffusers / transformers / torch on the oracle side; OS, Accelerate, and Wan2.1 weights revision not recorded in source notes.
• Validate-the-oracle caveat: the parity is meaningful only if the fp32 diffusers reference is itself correctly configured (matched seed, scheduler, step count, fp32 throughout). Reproducers should pin and verify the reference configuration before trusting the cosine/PSNR figures.
• Known refactor of record: the UMT5 encoder linear was moved from a strided cache-hostile loop to an Accelerate BLAS matmul; per-layer cosine remained 1.0, confirming the optimization did not perturb parity.