HISTORICAL PORT RECORD · SOURCE REVIEW
The earlier Timeline Studio port record describes how two unsupported 5D GridSample nodes were rewritten for browser WebGPU and compared against fixed-input outputs. This page preserves that numerical evidence. Our September 2026 source review also identifies the limits of the record: its timing is historical, its audiovisual result was not reproduced here, and current frame checks do not establish convincing speech or identity.
The browser workflow is not one model call. JoyVASA converts audio into motion, while LivePortrait performs appearance extraction, motion extraction, lip retargeting, stitching, feature warping and SPADE image generation. Keeping those stages separate creates useful validation boundaries: audio-driver error, keypoint error, warping error and final generator error can be measured independently rather than diagnosed from one incorrect-looking face.
The opset-20 combined generator had 277 nodes, including 13 rank-5 convolutions and two rank-5 GridSample nodes. A Linux shared-library plugin from the source repository could not run in the browser, and ONNX Runtime Web's WebGPU execution provider could not directly execute the 5D samplers. Removing the depth dimension, treating it as an arbitrary batch or selecting the nearest slice would all change the function computed by the model.
A target depth coordinate lies between slices d0 and d1. The rewrite performs supported 4D bilinear GridSample on both planes, then interpolates the two results using the fractional depth coordinate alpha. Depth slices are arranged as batches so WebGPU can process them in parallel and then restored to the original spatial layout. A verified shared-XY structure in the sparse-motion grid removes redundant sampling. The rewritten graph grew from 277 to 316 nodes, but every operation could run in the browser.
The earlier port record reports a comparison of the original 512 × 512 generator and rewritten graph on fixed inputs: maximum absolute error 7.8976e-6, mean absolute error 1.3853e-7 and 99.9th-percentile error 1.7881e-6. It also records maximum ONNX-versus-PyTorch errors of 7.75e-6 for JoyVASA's HuBERT export and 6.44e-6 for its diffusion denoiser. These are preserved development measurements, not values rerun during the September 2026 review. They support the specific recorded numerical comparison; they do not establish quality across different portraits or speech.
The WebGPU workers configure matching ONNX Runtime Asyncify JavaScript and WASM files and create execution-provider sessions serially. The render worker retains the portrait appearance feature for reuse. Summing the JoyVASA and LivePortrait configured artifacts gives 738,879,580 bytes for preview, or 949,593,258 bytes for both resolution tiers: approximately 739 MB and 950 MB respectively, equivalent to 705 MiB and 906 MiB. These are file-size calculations, excluding the application, runtime and inference memory; they are not measured downloads. The model files are loaded from a pinned hosted revision.
The configured mixed-FP16 tiers produce 256-pixel preview or 512-pixel quality images. The earlier development note reports roughly 25–60 seconds per 512 × 512 neural frame on its test browser and GPU. It does not name that browser/GPU configuration or attach a complete timing log. We did not rerun that benchmark in this review and cannot transfer its speed to the current device or the 256-pixel tier.
The current implementation renders sparse neural keyframes and copies the nearest available image into each encoded timestamp. An eight-fps WebM therefore does not mean eight neural images per second or motion interpolation. Downloads, session initialization, JoyVASA motion generation and encoding add costs outside the historical per-frame rendering figure.
The development note reports that model-part loading, WebGPU setup, 50-step JoyVASA motion, LivePortrait rendering, WebM encoding, browser playback and visual-track replacement completed. It describes a playable 3.56-second WebM, changing mouth states, head and eye movement, and no browser console errors after completion. The original voiceover, output video and full run log were not supplied with that note and were not revalidated in this assessment. The report establishes what was recorded during development; it is not a new synchronized audiovisual example a reader can inspect here.
A fresh practical acceptance run should preserve the exact portrait, speech, intermediate video and final export with sound. The current avatar encoder adds only a video track to the intermediate WebM. Audio must be verified after the timeline export rather than inferred from a successful video encode.
A useful graph-port record should identify tensor names, rank, axis order, numeric type and output ownership. Batch, channel, depth, height and width have distinct meanings in the feature volume; reshapes and transposes must preserve them. The preserved port note supplies baseline graph hashes and fixed-input error statistics, while the smoke script documents a small renderer path.
That smoke script runs WASM and requests fixed lip ratios; it is not a JoyVASA audio-driven browser benchmark. Dynamic-shape coverage, mixed-precision regression, current WebGPU speed and audiovisual alignment would need separate tests and retained artifacts. We did not execute those tests in this review and do not claim that the recorded fixed-input comparison covers them.
The implemented guards inspect sampled finite values, extreme magnitude, near-flat variance and distance from a reference frame. The worker retries a rejected keyframe once and can repeat the previous valid frame if both attempts fail. This can keep an obviously damaged image out of the encoded sequence, but a held frame is still a visible loss of motion.
No phoneme alignment, speech-to-mouth score or face-identity metric is implemented by those numerical guards. The current nearest-keyframe reuse also provides no generated intermediate motion whose identity preservation has been measured. To judge a performance, retain a final export with the actual voiceover and review lip closure, teeth, eyes, jaw, start/end timing and any fallback holds. The talking-portrait assessment proposes a one-second preview and a stopping budget before a more expensive run; that perceptual trial has not yet been performed for this review.
ENGINEERING RECORD / NUMERICAL VALIDATION
The port record preserves the pinned graph hashes and numerical comparison used to accept the WebGPU rewrite. Those measurements are stronger evidence than a single favorable portrait frame, while still leaving speed and generalization as explicit limits.
SEPTEMBER 7, 2026 · SOURCE REVIEW AND CORRECTIONS
We checked the implementation at commit 68980d1 to narrow claims that the earlier article could not support. This revision records source findings and model-size arithmetic. The related portrait assessment now adds a reproducible non-neural encoder experiment with three test cards. Neither page includes new neural inference, a dynamic portrait-quality test or a current neural-speed benchmark.
The structured source review records these boundaries and the proposed one-second trial. The SHA-256 manifest identifies every preserved source file and excerpt; the method and file guide explains their scope. A current final video with sound, perceptual review and complete timing remains future evidence.
REFERENCE
Browser runtimes cannot load the original Linux shared library, so the computation must use portable supported operators or a browser-specific implementation.
The earlier fixed-input comparison records errors around 1e-6 to 1e-7. Those measurements were not rerun in this review and do not establish equivalence for every input.
No current speed benchmark is established here. The older port note reports 25–60 seconds per 512-pixel neural frame; current eight-fps encoding reuses sparse keyframes and is not real-time neural generation.
A fixed revision keeps cached graph files, tensor contracts and runtime assumptions synchronized across sessions.
