Use ShowText cached output for LLM prompt chains

2026-05-28 22:58:18 +08:00
parent 4d8cc18219
commit 5d359d3a54
4 changed files with 261 additions and 3 deletions
@@ -187,7 +187,20 @@ class TextResolutionContext:

    def is_llm_template_node(self, node: NodeRecord) -> bool:
        lowered = node.class_type.lower()
-        markers = ("gemini", "chatgpt", "claude", "llm", "language_model")
+        markers = (
+            "gemini",
+            "openai",
+            "chatgpt",
+            "claude",
+            "llm",
+            "vlm",
+            "language_model",
+            "image caption",
+            "caption",
+            "prompt generator",
+            "promptgenerator",
+            "tagger",
+        )
        return any(marker in lowered for marker in markers)

    def record_unresolved(
@@ -34,7 +34,7 @@ class StringFunctionAdapter:

 class ShowTextAdapter:
    def matches(self, node: NodeRecord) -> bool:
-        return node.class_type == "ShowText|pysssss"
+        return _is_show_text_node(node)

    def resolve(
        self,
@@ -50,6 +50,10 @@ class ShowTextAdapter:
        if cached_text:
            text, source = cached_text
            confidence = 0.75 if source == "workflow widget cache fallback" else 1.0
+            context.trace.add(
+                "TEXT_TRACE",
+                f"{node.class_type} {node.node_id}: using cached ShowText text from {source}",
+            )
            if source == "workflow widget cache fallback":
                context.trace.add(
                    "TEXT_TRACE",
@@ -69,6 +73,23 @@ class ShowTextAdapter:
                )
            ]

+        upstream_llm = self._upstream_llm_runtime_node(node, context)
+        if upstream_llm is not None:
+            context.trace.add(
+                "TEXT_TRACE",
+                (
+                    f"{upstream_llm.class_type} {upstream_llm.node_id}: low confidence "
+                    "LLM template input skipped, no cached generated output found"
+                ),
+            )
+            context.record_unresolved(
+                node=node,
+                field="text",
+                role=role,
+                reason="ShowText cache missing and upstream is LLM runtime output not embedded",
+            )
+            return []
+
        context.trace.add(
            "TEXT_TRACE",
            (
@@ -94,3 +115,21 @@ class ShowTextAdapter:
            if workflow:
                return workflow
        return None
+
+    def _upstream_llm_runtime_node(
+        self,
+        node: NodeRecord,
+        context: TextResolutionContext,
+    ) -> NodeRecord | None:
+        target = context.graph.link_target(node.inputs.get("text"))
+        if target is None:
+            return None
+        upstream = context.graph.get_node(target[0])
+        if upstream is not None and context.is_llm_template_node(upstream):
+            return upstream
+        return None
+
+
+def _is_show_text_node(node: NodeRecord) -> bool:
+    lowered = node.class_type.lower().replace("_", "").replace(" ", "")
+    return lowered == "showtext" or lowered.startswith("showtext|")
@@ -64,5 +64,18 @@ class GenericTextAdapter:


 def context_marker_is_llm(lowered_class_type: str) -> bool:
-    markers = ("gemini", "chatgpt", "claude", "llm", "language_model")
+    markers = (
+        "gemini",
+        "openai",
+        "chatgpt",
+        "claude",
+        "llm",
+        "vlm",
+        "language_model",
+        "image caption",
+        "caption",
+        "prompt generator",
+        "promptgenerator",
+        "tagger",
+    )
    return any(marker in lowered_class_type for marker in markers)
@@ -115,6 +115,93 @@ def _upscale_workflow(reader_image="003.png"):
    }


+def _llm_showtext_upscale_prompt(show_class="ShowText|pysssss", include_cache=True):
+    return {
+        "22": {
+            "class_type": "FaceDetailer",
+            "inputs": {"image": ["24", 0], "positive": ["26", 0], "negative": ["27", 0]},
+        },
+        "24": {
+            "class_type": "UltimateSDUpscale",
+            "inputs": {
+                "image": ["40", 0],
+                "positive": ["26", 0],
+                "negative": ["27", 0],
+                "seed": 222,
+                "steps": 9,
+                "cfg": 1.5,
+                "sampler_name": "Euler",
+                "scheduler": "normal",
+            },
+        },
+        "26": {"class_type": "CLIPTextEncode", "inputs": {"text": ["37", 0]}},
+        "27": {"class_type": "CLIPTextEncode", "inputs": {"text": "bad hands, low quality"}},
+        "37": {"class_type": show_class, "inputs": {"text": ["90", 0]}},
+        "40": {"class_type": "LoadImage", "inputs": {"image": "source.png"}},
+        "90": {
+            "class_type": "GeminiChatNode",
+            "inputs": {
+                "system_instruction": "system instruction should not enter prompt",
+                "prompt": "Gemini prompt template should not enter prompt",
+                "api_key": "secret-key",
+                "model": "gemini-test-model",
+            },
+        },
+        "91": {"class_type": "ShowText|pysssss", "inputs": {"text": ["90", 0]}},
+        "99": {"class_type": "PreviewImage", "inputs": {"images": ["22", 0]}},
+    }
+
+
+def _llm_showtext_workflow(show_class="ShowText|pysssss", cached_text="cached gemini positive prompt"):
+    widgets_values = [cached_text] if cached_text else []
+    return {
+        "nodes": [
+            {
+                "id": 37,
+                "type": show_class,
+                "inputs": {"text": ["90", 0]},
+                "widgets_values": widgets_values,
+            },
+            {
+                "id": 90,
+                "type": "GeminiChatNode",
+                "inputs": {
+                    "system_instruction": "system instruction should not enter prompt",
+                    "prompt": "Gemini prompt template should not enter prompt",
+                    "api_key": "secret-key",
+                    "model": "gemini-test-model",
+                },
+                "widgets_values": [
+                    "Gemini prompt template should not enter prompt",
+                    "system instruction should not enter prompt",
+                    "secret-key",
+                    "gemini-test-model",
+                ],
+            },
+            {
+                "id": 91,
+                "type": "ShowText|pysssss",
+                "inputs": {"text": ["90", 0]},
+                "widgets_values": ["wrong unused prompt"],
+            },
+        ]
+    }
+
+
+def _bundle_for_prompt(prompt, workflow=None, filename="sample.png", width=1200, height=1600):
+    from smart_metadata_reader.models import MetadataBundle
+
+    return MetadataBundle(
+        filename=filename,
+        width=width,
+        height=height,
+        prompt_raw=json.dumps(prompt),
+        workflow_raw=json.dumps(workflow) if workflow is not None else None,
+        parameters_raw=None,
+        source_format="ComfyUI prompt/workflow",
+    )
+
+
 def test_read_metadata_extracts_png_text_chunks_and_dimensions(tmp_path):
    from smart_metadata_reader.metadata_reader import read_metadata

@@ -609,3 +696,109 @@ def test_nested_smart_metadata_reader_respects_actual_connected_output_role(tmp_

    assert result.positive == "original negative prompt"
    assert "SmartMetadataReader output negative used in positive chain" in result.debug_trace
+
+
+def test_showtext_cache_from_llm_chain_is_used_without_reading_llm_template():
+    from smart_metadata_reader.metadata_reader import parse_metadata_bundle
+
+    result = parse_metadata_bundle(
+        _bundle_for_prompt(
+            _llm_showtext_upscale_prompt(),
+            workflow=_llm_showtext_workflow(),
+        )
+    )
+
+    assert result.positive == "cached gemini positive prompt"
+    assert result.negative == "bad hands, low quality"
+    assert "Gemini prompt template" not in result.positive
+    assert "system instruction" not in result.positive
+    assert "secret-key" not in result.positive
+    assert result.status_message != "FAILED"
+    assert "using cached ShowText text" in result.debug_trace
+    assert "wrong unused prompt" not in result.positive
+
+
+def test_showtext_cache_compatible_class_from_llm_chain_is_used():
+    from smart_metadata_reader.metadata_reader import parse_metadata_bundle
+
+    result = parse_metadata_bundle(
+        _bundle_for_prompt(
+            _llm_showtext_upscale_prompt(show_class="ShowText"),
+            workflow=_llm_showtext_workflow(show_class="ShowText"),
+        )
+    )
+
+    assert result.positive == "cached gemini positive prompt"
+    assert "Gemini prompt template" not in result.positive
+
+
+def test_showtext_cache_is_single_segment_inside_string_function_chain():
+    from smart_metadata_reader.metadata_reader import parse_metadata_bundle
+
+    prompt = _llm_showtext_upscale_prompt()
+    prompt["26"]["inputs"]["text"] = ["39", 0]
+    prompt["39"] = {
+        "class_type": "StringFunction|pysssss",
+        "inputs": {
+            "action": "append",
+            "text_a": "manual prompt",
+            "text_b": ["37", 0],
+            "text_c": "preset prompt",
+        },
+    }
+
+    result = parse_metadata_bundle(
+        _bundle_for_prompt(prompt, workflow=_llm_showtext_workflow())
+    )
+
+    assert "manual prompt" in result.positive
+    assert "cached gemini positive prompt" in result.positive
+    assert "preset prompt" in result.positive
+    assert "Gemini prompt template" not in result.positive
+    assert "system instruction" not in result.positive
+    assert result.positive.index("manual prompt") < result.positive.index("cached gemini positive prompt")
+    assert result.positive.index("cached gemini positive prompt") < result.positive.index("preset prompt")
+
+
+def test_conditioning_combine_showtext_cache_does_not_override_other_branches():
+    from smart_metadata_reader.metadata_reader import parse_metadata_bundle
+
+    prompt = _llm_showtext_upscale_prompt()
+    prompt["24"]["inputs"]["positive"] = ["41", 0]
+    prompt["41"] = {
+        "class_type": "ConditioningCombine",
+        "inputs": {"conditioning_1": ["42", 0], "conditioning_2": ["43", 0]},
+    }
+    prompt["42"] = {"class_type": "CLIPTextEncode", "inputs": {"text": "branch A manual prompt"}}
+    prompt["43"] = {"class_type": "CLIPTextEncode", "inputs": {"text": ["37", 0]}}
+
+    result = parse_metadata_bundle(
+        _bundle_for_prompt(prompt, workflow=_llm_showtext_workflow())
+    )
+
+    assert "branch A manual prompt" in result.positive
+    assert "cached gemini positive prompt" in result.positive
+    assert "Gemini prompt template" not in result.positive
+
+
+def test_showtext_missing_cache_with_llm_upstream_returns_partial_unresolved():
+    from smart_metadata_reader.metadata_reader import parse_metadata_bundle
+
+    result = parse_metadata_bundle(
+        _bundle_for_prompt(
+            _llm_showtext_upscale_prompt(include_cache=False),
+            workflow=_llm_showtext_workflow(cached_text=""),
+        )
+    )
+
+    assert result.positive == ""
+    assert "Gemini prompt template" not in result.positive
+    assert "system instruction" not in result.positive
+    assert result.status_message == "PARTIAL"
+    assert result.confidence <= 0.6
+    assert result.partial_result["unresolved"][0]["class_type"] == "ShowText|pysssss"
+    assert (
+        result.partial_result["unresolved"][0]["reason"]
+        == "ShowText cache missing and upstream is LLM runtime output not embedded"
+    )
+    assert "Unresolved:" in result.setting