Use the general checkpoint for uploads; drop "MLP" from the title

Uploaded images and videos are now reconstructed with
DVSM-TTB-p8-general-960p. The gallery examples were baked with
DVSM-TTB-p8-DL3DV-960p, so both models are loaded and a live session
renders each artifact with the model named in its hf_variant.

APP_MODEL_DIR now points at a directory with one sub-directory per
variant (the nvidia/DVSM layout) instead of a single model directory.

Wording: the subtitle reads "Test-time Baking ..." and the tagline says
"TTB bakes the scene ...".

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>

README.md

@@ -16,7 +16,7 @@ license: other
 Upload images / a short orbit video → π³ pose estimation + TTB-DVSM
 reconstruction (one GPU call) → a time-boxed interactive viser viewer.
 
-Fixed model: `DVSM-TTB-p8-DL3DV-960p` (`nvidia/DVSM`).
+Models (`nvidia/DVSM`): uploads are reconstructed with `DVSM-TTB-p8-general-960p`; the gallery examples were baked with `DVSM-TTB-p8-DL3DV-960p`.
 
 ## ZeroGPU notes
 Select **ZeroGPU** hardware in the Space Settings (it is not set via this

app.py

@@ -17,7 +17,9 @@ Why this shape (ZeroGPU constraints, see README):
     "Start / Continue" to fork a new worker and rebuild the viewer from the
     already-computed artifact on disk (no re-reconstruction).
 
-Fixed model: DVSM-TTB-p8-DL3DV-960p (nvidia/DVSM).
+Models (nvidia/DVSM): uploads are reconstructed with DVSM-TTB-p8-general-960p; the
+gallery examples were baked with DVSM-TTB-p8-DL3DV-960p. A live session renders an
+artifact with the model that baked it.
 """
 from __future__ import annotations
 
@@ -61,7 +63,11 @@ from viser_app.zerogpu_compat import repair_torch_cuda_state
 # ---------------------------------------------------------------------------
 # Config
 # ---------------------------------------------------------------------------
-HF_VARIANT = "DVSM-TTB-p8-DL3DV-960p"
+# Uploaded images / videos are reconstructed with the general model. The gallery
+# examples were baked with the DL3DV model. An artifact records the variant that
+# baked it ("hf_variant") and has to be rendered with that same model.
+UPLOAD_VARIANT = "DVSM-TTB-p8-general-960p"
+EXAMPLE_VARIANT = "DVSM-TTB-p8-DL3DV-960p"
 # How long one held-GPU interactive session lasts before it must be renewed.
 # Bounded by the per-call duration cap and the user's daily quota
 # (anon 2min / free 5min / PRO 40min). Override via env for PRO accounts.
@@ -102,18 +108,25 @@ EXAMPLE_SCENES = _discover_examples()
 # Module-level model load: download + keep on CPU (ZeroGPU convention).
 # Real `.cuda()` happens inside @spaces.GPU workers.
 # ---------------------------------------------------------------------------
-print(f"[app] loading {HF_VARIANT} + π³ to CPU ...", flush=True)
+print(f"[app] loading {UPLOAD_VARIANT}, {EXAMPLE_VARIANT} + π³ to CPU ...", flush=True)
 from dvsm.models.DVSM import DecoderViewSynthesisModel
 from viser_app.pi3_estimator import load_pi3_model
 
-# APP_MODEL_DIR: load the DVSM model from a local HF-format directory instead of
-# downloading from the Hub (faster / offline; used for local GPU-node testing).
+# APP_MODEL_DIR: load the DVSM models from a local directory instead of the Hub
+# (faster / offline; used for local GPU-node testing). It holds one HF-format
+# sub-directory per variant, i.e. the layout of the nvidia/DVSM repo.
 _LOCAL_MODEL_DIR = os.environ.get("APP_MODEL_DIR")
-if _LOCAL_MODEL_DIR:
-    print(f"[app] loading DVSM from local dir {_LOCAL_MODEL_DIR}", flush=True)
-    DVSM_MODEL_CPU = DecoderViewSynthesisModel.from_pretrained(_LOCAL_MODEL_DIR).eval()
-else:
-    DVSM_MODEL_CPU = DecoderViewSynthesisModel.from_hf_pretrained(HF_VARIANT).eval()
+
+
+def _load_dvsm(variant):
+    if _LOCAL_MODEL_DIR:
+        path = Path(_LOCAL_MODEL_DIR) / variant
+        print(f"[app] loading {variant} from local dir {path}", flush=True)
+        return DecoderViewSynthesisModel.from_pretrained(str(path)).eval()
+    return DecoderViewSynthesisModel.from_hf_pretrained(variant).eval()
+
+
+DVSM_MODELS_CPU = {v: _load_dvsm(v) for v in (UPLOAD_VARIANT, EXAMPLE_VARIANT)}
 PI3_MODEL_CPU = load_pi3_model(flavor="pi3x", device="cpu").eval()
 print("[app] models ready (CPU)", flush=True)
 
@@ -167,11 +180,11 @@ def reconstruct_gpu(img_dir: str, target_h: int, target_w: int, max_views: int)
     # Reuse the module-level CPU model (moved to cuda inside reconstruct()).
     renderer, batch, ishape, c2w_r, c2w_t, intr = reconstruct(
         ckpt_root=None, ckpt_name="", scene_data=scene_data, cmd_lst=[],
-        hf_variant=HF_VARIANT, model=DVSM_MODEL_CPU,
+        hf_variant=UPLOAD_VARIANT, model=DVSM_MODELS_CPU[UPLOAD_VARIANT],
     )
     out_path = ARTIFACTS_DIR / f"{int(time.time() * 1000)}_{uuid.uuid4().hex[:8]}.recon.pt"
     save_artifact(renderer, batch, ishape, c2w_r, c2w_t, intr, out_path,
-                  hf_variant=HF_VARIANT)
+                  hf_variant=UPLOAD_VARIANT)
     return str(out_path)
 
 
@@ -206,8 +219,13 @@ def _live_session_worker(artifact_path: str):
 
     repair_torch_cuda_state()  # torch.compile vs ZeroGPU, see zerogpu_compat
     yield ("prep", 0.15, "Loading scene…")
-    model = DVSM_MODEL_CPU.cuda().eval()
     artifact = torch.load(artifact_path, weights_only=False, map_location="cpu")
+    # Render with the model that baked this artifact (uploads vs gallery examples).
+    variant = artifact.get("hf_variant") or EXAMPLE_VARIANT
+    if variant not in DVSM_MODELS_CPU:
+        raise RuntimeError(f"artifact was baked with {variant}, which this app does not load")
+    print(f"[session] rendering with {variant}", flush=True)
+    model = DVSM_MODELS_CPU[variant].cuda().eval()
     renderer = build_renderer(model, artifact)
     scene_meta = artifact["scene_meta"]
 
@@ -605,14 +623,14 @@ with gr.Blocks(title="TTB-DVSM (ZeroGPU)", delete_cache=(3600, 3600)) as demo:
         "<h2 style='font-style:normal;font-weight:500;font-size:1.6em;"
         "margin-top:0;color:var(--body-text-color)'>"
         "<b style='color:#76B900'>T</b>est-<b style='color:#76B900'>t</b>ime "
-        "MLP <b style='color:#76B900'>B</b>aking for Efficient Rendering of "
+        "<b style='color:#76B900'>B</b>aking for Efficient Rendering of "
         "<b style='color:#76B900'>D</b>ecoder-only "
         "<b style='color:#76B900'>V</b>iew "
         "<b style='color:#76B900'>S</b>ynthesis "
         "<b style='color:#76B900'>M</b>odels"
         "</h2>"
         "<p style='color:var(--body-text-color);font-size:1.1em;margin-top:0.5em;line-height:1.6'>"
-        "TTB-DVSM bakes the scene into fast weights in a single forward pass, "
+        "TTB bakes the scene into fast weights in a single forward pass, "
         "and the result is browseable in the viser viewer."
         "</p>"
         "<div style='color:var(--body-text-color);font-size:1.1em;margin-top:0.8em;line-height:1.7'>"