Merge remote-tracking branch 'origin/image-generation' into diffusion-phase4-native

# Conflicts:
#	scripts/diffusion_bench.py
#	scripts/diffusion_quality.py
#	studio/backend/core/inference/diffusion.py
#	studio/backend/core/inference/diffusion_device.py
#	studio/backend/core/inference/diffusion_families.py
#	studio/backend/core/inference/diffusion_memory.py
#	studio/backend/core/inference/diffusion_precision.py
#	studio/backend/core/inference/diffusion_speed.py
#	studio/backend/models/inference.py
#	studio/backend/routes/inference.py
#	studio/backend/tests/test_diffusion_backend.py
#	studio/backend/tests/test_diffusion_device.py
#	studio/backend/tests/test_diffusion_memory.py
#	studio/backend/tests/test_diffusion_precision.py
#	studio/backend/tests/test_diffusion_speed.py

scripts/diffusion_bench.py

@@ -142,8 +142,10 @@ def _psnr(ref_png: Path, cand_png: Path) -> float:
     import numpy as np
     from PIL import Image
 
-    a = np.asarray(Image.open(ref_png).convert("RGB"), dtype = np.float64)
-    b = np.asarray(Image.open(cand_png).convert("RGB"), dtype = np.float64)
+    with Image.open(ref_png) as im_a:
+        a = np.asarray(im_a.convert("RGB"), dtype = np.float64)
+    with Image.open(cand_png) as im_b:
+        b = np.asarray(im_b.convert("RGB"), dtype = np.float64)
     if a.shape != b.shape:
         # Different geometry means the comparison is meaningless; report worst case.
         return 0.0
@@ -360,9 +362,14 @@ def _compare(args: argparse.Namespace) -> int:
             print("   refusing noisy comparison (pass --force-compare to override).", flush = True)
             return 2
 
-    # PSNR vs the stored reference image.
+    # PSNR vs the stored reference image. The baseline stores an absolute reference_png,
+    # which breaks if the baseline directory was copied/moved, so fall back to reference.png
+    # next to the baseline JSON. A still-missing reference is a failure below, not a silent
+    # pass -- otherwise the benchmark would report PASS having done no image comparison.
     ref_png = Path(baseline.get("accuracy", {}).get("reference_png", ""))
-    psnr = _psnr(ref_png, args._image_out) if ref_png.exists() else float("nan")
+    if not ref_png.is_file():
+        ref_png = baseline_path.parent / "reference.png"
+    psnr = _psnr(ref_png, args._image_out) if ref_png.is_file() else float("nan")
 
     base_gen = baseline.get("generate", {})
     cur_gen = metrics["generate"]
@@ -394,7 +401,9 @@ def _compare(args: argparse.Namespace) -> int:
         )
     if base_peak and cur_peak and vram_reg > args.max_vram_regression:
         failures.append(f"peak VRAM +{vram_reg * 100:.1f}% > {args.max_vram_regression * 100:.0f}%")
-    if not math.isnan(psnr) and psnr < args.min_psnr:
+    if math.isnan(psnr):
+        failures.append("PSNR reference image missing; cannot verify output quality")
+    elif psnr < args.min_psnr:
         failures.append(f"PSNR {psnr:.2f}dB < {args.min_psnr:.1f}dB (output changed)")
 
     if failures:

scripts/diffusion_quality.py

@@ -186,6 +186,18 @@ def _wait_for_load(backend: Any, timeout_s: int = 3600) -> None:
 
 
 def _hf_file_size_mib(repo: str, filename: str) -> Optional[int]:
+    # A local model dir / file: stat it directly. The Hub lookup below returns None for
+    # a local path, which would drop every candidate from _recommend (file_size_mib None).
+    try:
+        local = Path(repo).expanduser()
+        if local.is_dir():
+            f = local / filename
+            if f.is_file():
+                return int(f.stat().st_size // (1024 * 1024))
+        elif local.is_file():
+            return int(local.stat().st_size // (1024 * 1024))
+    except Exception:
+        pass
     try:
         from huggingface_hub import HfApi
         info = HfApi().model_info(repo, files_metadata = True, token = os.environ.get("HF_TOKEN"))
@@ -268,6 +280,11 @@ def _compare(
             clip_sim.append(clip.image_similarity(img, ref))
 
     def _mean(xs: list[float]) -> Optional[float]:
+        # Preserve +inf: an identical render (reference vs itself, or a lossless
+        # quant/offload) scores PSNR=inf, which is exactly the case this harness
+        # verifies; dropping it as non-finite would print "-" instead of "inf".
+        if xs and any(x == math.inf for x in xs):
+            return math.inf
         finite = [x for x in xs if math.isfinite(x)]
         return round(sum(finite) / len(finite), 4) if finite else None
 

scripts/scan_packages.py

@@ -1208,9 +1208,10 @@ def check_js_file(content: str, filename: str, package: str) -> list[Finding]:
                 HIGH,
                 package,
                 filename,
-                f"Python wheel ships large ({len(content) // 1024} KB) JS bundle "
-                "(uncommon; manually review)",
-                "",
+                # Size stays in evidence, not the check label, so the baseline key
+                # does not drift when a wheel's bundle grows by a few KB.
+                "Python wheel ships large JS bundle (uncommon; manually review)",
+                f"{len(content) // 1024} KB JS bundle",
             )
         )
     return findings

scripts/scan_packages_baseline.json

@@ -1181,7 +1181,7 @@
     {
       "package": "tensorboard",
       "file": "tensorboard/plugins/projector/tf_projector_plugin/projector_binary.js",
-      "check": "Python wheel ships large (1918 KB) JS bundle (uncommon; manually review)",
+      "check": "Python wheel ships large JS bundle (uncommon; manually review)",
       "severity": "HIGH",
       "evidence": ""
     },