Judge: add seed (reproducible output) + interrupt support

New 'seed' input seeds sampling (torch.manual_seed before generation) so a fixed seed
reproduces the same output — no reroll when re-queueing while tweaking downstream nodes;
bump it to reroll (named 'seed' so the frontend adds control_after_generate). Generation
now honors ComfyUI cancel: a StoppingCriteria halts model.generate() promptly and
_check_interrupt() raises between passes. Workflows updated with the seed widget.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

nodes/qwen_judge.py

@@ -453,8 +453,33 @@ def _apply_template(processor, messages, think=True):
         return _format_chatml_qwenvl(messages)
 
 
+def _check_interrupt():
+    """Raise if the user hit ComfyUI's cancel — aborts the node between passes."""
+    try:
+        import comfy.model_management as mm
+        mm.throw_exception_if_processing_interrupted()
+    except ImportError:
+        pass
+
+
+def _interrupt_stopping():
+    """StoppingCriteria that halts generate() promptly when ComfyUI is interrupted."""
+    try:
+        import comfy.model_management as mm
+        from transformers import StoppingCriteria, StoppingCriteriaList
+
+        class _Interrupt(StoppingCriteria):
+            def __call__(self, input_ids, scores, **kw):
+                return mm.processing_interrupted()
+
+        return StoppingCriteriaList([_Interrupt()])
+    except Exception:
+        return None
+
+
 def _generate_from_messages(model, processor, messages, images, max_new_tokens, temperature, think=True):
     """Template + forward pass for a chat-message list; returns the decoded string."""
+    _check_interrupt()
     text = _apply_template(processor, messages, think)
     inputs = processor(text=[text], images=images, return_tensors="pt")
     inputs = inputs.to(model.device)
@@ -464,6 +489,9 @@ def _generate_from_messages(model, processor, messages, images, max_new_tokens,
         gen_kwargs.update(do_sample=True, temperature=float(temperature))
     else:
         gen_kwargs.update(do_sample=False)
+    sc = _interrupt_stopping()
+    if sc is not None:
+        gen_kwargs["stopping_criteria"] = sc
 
     with torch.inference_mode():
         out = model.generate(**inputs, **gen_kwargs)
@@ -821,6 +849,9 @@ class QwenVLImageJudge:
                 "precision": (["bf16", "fp8", "nf4"], {"default": "bf16"}),
                 "max_new_tokens": ("INT", {"default": 3072, "min": 64, "max": 8192}),
                 "temperature": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.5, "step": 0.05}),
+                # Seeds sampling so a fixed seed reproduces the same output (no reroll when
+                # you re-queue while tweaking downstream nodes). Bump it to reroll.
+                "seed": ("INT", {"default": 0, "min": 0, "max": 0xFFFFFFFFFFFFFFFF}),
                 "swap_eval": ("BOOLEAN", {"default": True}),
                 # Reasoning models (Qwen3.5/3.6) judge verdicts FAR better with thinking on
                 # (off -> they rubber-stamp 'match'). Costs more tokens; raise max_new_tokens.
@@ -849,7 +880,7 @@ class QwenVLImageJudge:
         }
 
     def judge(self, reference_image, mode, model_path, precision,
-              max_new_tokens, temperature, swap_eval, profile="general",
+              max_new_tokens, temperature, swap_eval, seed=0, profile="general",
               enable_thinking=True, json_output=False, model_select=MANUAL_CHOICE,
               generated_image=None, keep_loaded=True, auto_download=True,
               report_dir="", run_tag="", axes="", reference_description="",
@@ -888,6 +919,12 @@ class QwenVLImageJudge:
         ref_pil = _tensor_to_pil(reference_image)
         model, processor = _load_model(resolved_path, eff_precision)
 
+        # Seed sampling so a fixed seed reproduces the same output; check for cancel.
+        torch.manual_seed(int(seed))
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(int(seed))
+        _check_interrupt()
+
         if mode == "chat":
             gen_pil = _tensor_to_pil(generated_image) if generated_image is not None else None
             return self._chat(model, processor, ref_pil, gen_pil, system_prompt, user_prompt,