Drop the Advanced panel from the demo UI

Removes the Seed and Eta controls. Both are still passed to the rollout script,
just fixed internally: eta stays at the deterministic-ODE default (0.0), and
each run draws a fresh random seed rather than exposing -1 / explicit-seed
semantics that only make sense if you are reproducing a run.

Behaviour is unchanged from the previous defaults. The NUM_VIDEOS rollouts still
diverge, since the sampler draws independent initial noise per batch element
regardless of the shared seed. Demo text updated to stop pointing at a panel
that no longer exists.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app2.py

@@ -250,13 +250,12 @@ step emits several image frames, so the final clip is longer than the step count
 *L1 sampler steps* controls the detail predictor, which integrates the flow-matching
 ODE and benefits from more steps (sharper frames, slower). *L2 sampler steps* controls
 the abstract predictor, which is **consistency-distilled** and so is designed to run in
-very few steps — raising it buys little. *Eta* (under Advanced) injects noise per step:
-0 is the deterministic ODE, higher values turn it into an SDE.
+very few steps — raising it buys little.
 
 **Randomness.** The {NUM_VIDEOS} rollouts share one context and one seed, but the
-solver draws independent initial noise per batch element, so the futures diverge.
-Leave the seed at -1 under Advanced for a fresh draw each run, or set it explicitly
-to reproduce a run exactly.
+solver draws independent initial noise per batch element, so the futures diverge. Each
+run draws a fresh seed, so generating again on the same clip gives you a new set of
+futures.
 
 **Note.** The clip must be long enough for the L2 look-back window; very short clips
 are rejected with an explicit error. Frame and step counts are kept modest to fit the
@@ -349,7 +348,7 @@ def _stream_rollout(cmd: list[str], say) -> str:
 
 @spaces.GPU(duration=300)
 def run_rollout(video_path: str, num_gen_steps: int, l1_steps: int, l2_steps: int,
-                seed: int, eta: float, progress=gr.Progress()):
+                progress=gr.Progress()):
     if video_path is None:
         raise gr.Error("Please upload a short mp4 clip first.")
 
@@ -357,10 +356,13 @@ def run_rollout(video_path: str, num_gen_steps: int, l1_steps: int, l2_steps: in
         progress(frac, desc=msg)
 
     try:
-        # The rollout script only seeds when seed > 0; within the batch each sample
-        # still draws its own noise, so one seed yields NUM_VIDEOS distinct rollouts.
-        # seed < 0 means "surprise me" — draw a fresh positive seed each run.
-        seed = random.randrange(1, 2**31 - 1) if int(seed) < 0 else max(1, int(seed))
+        # Seed and eta are no longer exposed in the UI. Every run draws a fresh seed,
+        # so repeated runs on the same clip show different futures; the rollout script
+        # only seeds when seed > 0. eta stays at the deterministic-ODE default — within
+        # a batch each sample still draws its own noise, so the NUM_VIDEOS rollouts
+        # diverge regardless.
+        seed = random.randrange(1, 2**31 - 1)
+        eta = DEFAULT_ETA
 
         job = Path(tempfile.gettempdir()) / f"orbis2_{uuid.uuid4().hex[:8]}"
 
@@ -446,13 +448,6 @@ def build_demo():
                                      label="L2 sampler steps",
                                      info="Abstract predictor — distilled, needs very few")
 
-                with gr.Accordion("Advanced", open=False):
-                    seed_in = gr.Number(value=-1, precision=0, label="Seed",
-                                        info="-1 draws a fresh seed each run")
-                    eta = gr.Slider(0.0, 1.0, value=DEFAULT_ETA, step=0.05,
-                                    label="Eta (stochasticity)",
-                                    info="0 = deterministic ODE; >0 adds noise per step")
-
                 btn = gr.Button("Generate rollouts", variant="primary")
 
             with gr.Column(scale=1):
@@ -461,7 +456,7 @@ def build_demo():
 
         btn.click(
             run_rollout,
-            inputs=[inp, n_steps_gen, l1_steps, l2_steps, seed_in, eta],
+            inputs=[inp, n_steps_gen, l1_steps, l2_steps],
             outputs=outs,
             concurrency_limit=1,
         )