cross-theme warning, crossfade, and a swap time bound to the clip

app.py

@@ -214,6 +214,11 @@ WORLDS = {f"world {w['idx']} · {w['prompt'][:70]}": w["idx"] for w in INFO["wor
 WORLD_LABELS = list(WORLDS)
 PROMPT_LABELS = list(PROMPTS)
 W = {i: lbl for lbl, i in WORLDS.items()}   # by world id, for the default
+# Themes are what the local UI uses to warn about a steer that will smear. The four
+# shipped worlds are named by their own bank prompt, so the world id is a prompt id.
+from wanstreamer.serve.conditioning import theme_of  # noqa: E402
+PROMPT_THEME = {p["idx"]: p["theme"] for p in INFO["prompts"]}
+WORLD_THEME = {w["idx"]: theme_of(w["idx"]) for w in INFO["worlds"]}
 print(f"[boot] ready: step {INFO['step']}, {len(PROMPTS)} prompts, "
       f"{len(WORLDS)} worlds", flush=True)
 
@@ -245,7 +250,8 @@ def write_mp4(frames):
     return path
 
 
-def _duration(world, steer_to, steer_text, steer_at, seconds, seed, *args, **kwargs):
+def _duration(world, steer_to, steer_text, steer_at, crossfade, seconds, seed,
+              *args, **kwargs):
     # Measured: ~12 s to open a world (load, prime the cache, decode its 81 frames)
     # and ~1.02 s per block, a block being 0.75 s of video. Declared tight on purpose:
     # ZeroGPU compares the request against the visitor's remaining quota, not the
@@ -255,7 +261,7 @@ def _duration(world, steer_to, steer_text, steer_at, seconds, seed, *args, **kwa
 
 @spaces.GPU(duration=_duration)
 def run(world: str, steer_to: str, steer_text: str, steer_at: float,
-        seconds: float, seed: int):
+        crossfade: int, seconds: float, seed: int):
     """Stream video from a cached world, optionally swapping the prompt mid-stream.
 
     Args:
@@ -263,6 +269,7 @@ def run(world: str, steer_to: str, steer_text: str, steer_at: float,
         steer_to: a prompt from the 96-prompt bank to swap to, or the no-steer option.
         steer_text: free text to swap to instead; takes precedence over steer_to.
         steer_at: seconds into the clip at which to swap the conditioning.
+        crossfade: blocks to interpolate the conditioning over; 0 swaps instantly.
         seconds: how much video to generate, at 16 fps.
         seed: RNG seed for the block sampler.
 
@@ -270,7 +277,10 @@ def run(world: str, steer_to: str, steer_text: str, steer_at: float,
         (preview frame, HUD line, finished mp4) — the mp4 only on the last yield.
     """
     total_frames = int(float(seconds) * FPS)
-    steer_frame = int(float(steer_at) * FPS)
+    # A swap at or past the end of the clip would never fire and the run would look
+    # like the control did nothing. Keep it at least two seconds from the end so the
+    # result of the swap is actually visible.
+    steer_frame = int(min(float(steer_at), max(1.0, float(seconds) - 2.0)) * FPS)
 
     text = (steer_text or "").strip()
     if text:
@@ -319,7 +329,7 @@ def run(world: str, steer_to: str, steer_text: str, steer_at: float,
                                        cv2.IMREAD_COLOR)[:, :, ::-1])
 
             if pending_steer and len(frames) >= steer_frame:
-                engine.steer(**steer_with)
+                engine.steer(crossfade=int(crossfade), **steer_with)
                 pending_steer = False
 
             now = time.perf_counter()
@@ -357,6 +367,31 @@ Running on shared ZeroGPU hardware, so about 0.7x real time here; a dedicated GP
 does about 2.7x. [Code]({GITHUB}) · [Weights](https://huggingface.co/{LIVEWAN_REPO})
 """
 
+def steer_warning(world, steer_to, steer_text):
+    """The same warning the local UI shows before a steer that will smear.
+
+    Cross-theme steers ask the model to reconcile a K/V cache full of one kind of
+    scene with a prompt describing another, and it smears rather than resolving.
+    Free text gets no theme, so it cannot be checked.
+    """
+    if (steer_text or "").strip() or not steer_to or steer_to == NO_STEER:
+        return gr.update(visible=False)
+    wt = WORLD_THEME.get(WORLDS[world])
+    pt = PROMPT_THEME.get(PROMPTS[steer_to])
+    if not wt or not pt or wt == pt:
+        return gr.update(visible=False)
+    return gr.update(visible=True, value=(
+        f"⚠️ {wt.lower()} to {pt.lower()} is a jump across themes, so expect the "
+        f"picture to smear rather than resolve. Raise the crossfade to soften it, "
+        f"or steer somewhere closer to where you started."))
+
+
+def bound_swap(seconds, steer_at):
+    """A swap must land far enough before the end to be worth watching."""
+    hi = max(1.0, float(seconds) - 2.0)
+    return gr.update(maximum=hi, value=min(float(steer_at), hi))
+
+
 with gr.Blocks(title="LiveWan") as demo:
     with gr.Column(elem_id="col-container"):
         gr.Markdown(INTRO)
@@ -371,10 +406,15 @@ with gr.Blocks(title="LiveWan") as demo:
                     label="Or steer to your own text", lines=2,
                     placeholder="A lighthouse in a storm",
                     info="Encoded with umt5-xxl. Overrides the selection above.")
-                steer_at = gr.Slider(1, 25, value=6, step=0.5,
-                                     label="Swap at (seconds in)")
+                warning = gr.Markdown(visible=False)
                 seconds = gr.Slider(5, 30, value=15, step=1,
                                     label="Generate (seconds of video)")
+                steer_at = gr.Slider(1, 13, value=6, step=0.5,
+                                     label="Swap at (seconds in)")
+                crossfade = gr.Slider(0, 12, value=0, step=1,
+                                      label="Crossfade (blocks)",
+                                      info="0 swaps instantly. Blending over a few "
+                                           "blocks softens a big jump.")
                 seed = gr.Number(value=0, precision=0, label="Seed")
                 run_btn = gr.Button("Stream", variant="primary", size="lg")
 
@@ -386,7 +426,13 @@ with gr.Blocks(title="LiveWan") as demo:
                 video = gr.Video(label="The clip, at 16 fps", autoplay=True,
                                  height=368)
 
-    run_btn.click(run, inputs=[world, steer_to, steer_text, steer_at, seconds, seed],
+    for control in (world, steer_to, steer_text):
+        control.change(steer_warning, [world, steer_to, steer_text], warning)
+    seconds.change(bound_swap, [seconds, steer_at], steer_at)
+
+    run_btn.click(run,
+                  inputs=[world, steer_to, steer_text, steer_at, crossfade,
+                          seconds, seed],
                   outputs=[preview, stats, video], concurrency_limit=1)
 
 demo.queue(max_size=12).launch(theme=gr.themes.Citrus(), css=CSS, mcp_server=True)