JAM

musical accuracy research idle saved connecting…

Shot ?

Inputs

Drop a video or browse
Drop a video or browse
Drop a video or browse
All three are the same performance exported differently. Conditions B, C and D compare them head to head.
Drop an image or browse
Drop audio or browse
Drop the song's JSON folder here, or click to choose files
checking…
folder path
Set by the drop above. beats.json, guitar_tab_v2.json, lyrics.json, lyrics_sentiment.json — read by name, so keep the names.

Prompt

This is an instruction about the thing being measured. It appears in every condition including the floor — otherwise the ablation measures the wording, not the inputs.
H3 synthesises its own soundtrack unless told not to — the character then lip-syncs to invented vocals and the output can't be judged against the real track.
Worth sweeping on its own later — but hold it constant within a run.

Generate

stylemusic videoJSONwhat it answers
steps (4 as trained; 6 if soft)
Runs go to the backend chosen under Admin (). Cancel job stops the GPU; Stop only stops watching. Fetch last result re-downloads a finished job whose download or assembly failed — it bills nothing.

Status idle

Cells run one after another; each is one job.
No run yet.

Runs

Newest first. Blind hides the run name, condition, seed and model and shuffles the order, so a verdict cannot be steered by the label. Ratings are kept in this browser.

Admin

Default backend

Saved as the default. The Generate page never sends a backend of its own, so every run is comparable.

Comfy Cloud no key

RunComfy no token

Changing the SKU patches the saved deployment straight away. RunComfy bills GPU uptime, keep-warm included.
Creates a deployment on the SKU above, verifies the graph, and saves its ID here.
Reads the deployed graph on the default backend and reports what it is conditioned on. Spends nothing.