Size ZeroGPU duration from measured runs; keep /optimize api_name; document results

README.md

@@ -46,6 +46,19 @@ The **Learn2Splat** and **Adam (3DGS)** options run through the *same*
 minibatches and step budget — only the update rule differs. That is the
 like-for-like comparison the paper makes.
 
+## What it shows (measured on this Space)
+
+From the same SfM initialization (13.87 dB on the held-out views), 100
+refinement steps over 161 training views, 8 views per step, no densification:
+
+| optimizer | held-out PSNR | time |
+|---|---|---|
+| Learn2Splat (learned) | **23.91 dB** | 12.2 s |
+| Adam (3DGS baseline)  | 20.31 dB | 9.6 s |
+
+200 learned steps reach 24.34 dB in 21.8 s. Numbers move a little run to run —
+the per-step view minibatch is sampled at random.
+
 ## Implementation notes
 
 - Port of `demo.py --with-gui gradio` from the official

app.py

@@ -445,11 +445,15 @@ def head_pct(step: int, n_steps: int) -> str:
 # The run
 # --------------------------------------------------------------------------
 def _estimate_duration(method=METHOD_LEARNED, steps=DEFAULT_STEPS, *args, **kwargs):
+    """GPU seconds to reserve. Measured on the live Space (ZeroGPU / Blackwell):
+    ~0.11 s per step warm, plus ~13 s fixed (cold-worker warm-up, moving the
+    scene to VRAM, the held-out PSNR pass and the PLY export). A cold 30-step
+    run took 16.3 s end to end; this leaves ~1.4x headroom on top."""
     try:
-        s = int(steps)
+        n = int(steps)
     except Exception:
-        s = DEFAULT_STEPS
-    return int(min(280, 45 + 1.1 * s))
+        n = DEFAULT_STEPS
+    return int(min(180, 20 + 0.18 * n))
 
 
 @spaces.GPU(duration=_estimate_duration)
@@ -612,7 +616,7 @@ with gr.Blocks(title="Learn2Splat") as demo:
                 run_btn = gr.Button("Run optimization", variant="primary")
                 status_html = gr.HTML(
                     _msg("SfM-initialized — pick an optimizer, then <b>Run</b>."),
-                    elem_id="l2s-status",
+                    label="Run status", show_label=False, elem_id="l2s-status",
                 )
                 with gr.Accordion("Advanced settings", open=False):
                     view_sl = gr.Slider(
@@ -663,6 +667,7 @@ with gr.Blocks(title="Learn2Splat") as demo:
             # would defeat the point of watching the optimization converge.
             cache_examples=False,
             run_on_click=True,
+            api_name=False,  # keep /optimize for the Run button's handler
             label="Try",
         )