I built Avatar Video Workbench, an MIT-licensed control plane for reproducible avatar LoRA and LTX image-to-video experiments.
Demo Space:
GitHub:
The Space is text-only: it validates sanitized Vertex CustomJob reports and dataset manifests before sharing them publicly. It checks for cloud URIs, project IDs, service accounts, local paths, generated media references, model artifacts, tokens, and NSFW wording.
The repo currently covers:
a public-safe photo-to-character dataset route for authorized source images; LTX LoRA training CustomJob staging on Vertex AI;
LTX image-to-video CustomJob staging with optional LoRA weights;
backend metadata export for comparing scripted/cloud runs with ComfyUI-style workflows;
sanitized Vertex run reports and a publication scanner.
I would value feedback on:
which metadata fields are useful when comparing LTX or ComfyUI-style video runs;
whether the public report schema is missing runtime or cost fields;
common leak patterns the scanner should catch before maintainers publish examples.
No personal images, generated videos, model weights, or credentials are included.