[diffusion] CI: reset thresholds (#22854)

scripts/ci/utils/diffusion/comparison_configs.json

@@ -138,8 +138,11 @@
         {
             "id": "ltx2.3_twostage_ti2v_2gpus",
             "model": "Lightricks/LTX-2.3",
-            "task": "image-to-video",
+            "task": "text-image-to-video",
             "prompt": "The cat starts walking slowly towards the camera.",
+            "reference_image": true,
+            "width": 768,
+            "height": 512,
             "num_frames": 121,
             "seed": 42,
             "num_gpus": 2,

scripts/ci/utils/diffusion/run_comparison.py

@@ -959,7 +959,7 @@ def main():
 
     print(f"Loaded {len(config['cases'])} comparison case(s) from {args.config}")
 
-    run_comparison(
+    output_data = run_comparison(
         config=config,
         case_ids=args.case_ids,
         frameworks=args.frameworks,
@@ -968,6 +968,14 @@ def main():
         dry_run=args.dry_run,
     )
 
+    # Exit with non-zero if any case had an error
+    errors = [r for r in output_data.get("results", []) if r.get("error")]
+    if errors and not args.dry_run:
+        print(f"\n{len(errors)} case(s) had errors:")
+        for e in errors:
+            print(f"  {e['case_id']} ({e['framework']}): {e['error']}")
+        sys.exit(1)
+
 
 if __name__ == "__main__":
     main()