feat(api-nodes): add new Gemini model (Comfy-Org#10789)

bigcat88 · web-flow · commit 24fdb92edf2e · 2025-11-18T14:26:44.000-08:00
diff --git a/comfy_api_nodes/apis/gemini_api.py b/comfy_api_nodes/apis/gemini_api.py
@@ -1,22 +1,229 @@
-from typing import Optional
+from datetime import date
+from enum import Enum
+from typing import Any
 
-from comfy_api_nodes.apis import GeminiGenerationConfig, GeminiContent, GeminiSafetySetting, GeminiSystemInstructionContent, GeminiTool, GeminiVideoMetadata
-from pydantic import BaseModel
+from pydantic import BaseModel, Field
+
+
+class GeminiSafetyCategory(str, Enum):
+    HARM_CATEGORY_SEXUALLY_EXPLICIT = "HARM_CATEGORY_SEXUALLY_EXPLICIT"
+    HARM_CATEGORY_HATE_SPEECH = "HARM_CATEGORY_HATE_SPEECH"
+    HARM_CATEGORY_HARASSMENT = "HARM_CATEGORY_HARASSMENT"
+    HARM_CATEGORY_DANGEROUS_CONTENT = "HARM_CATEGORY_DANGEROUS_CONTENT"
+
+
+class GeminiSafetyThreshold(str, Enum):
+    OFF = "OFF"
+    BLOCK_NONE = "BLOCK_NONE"
+    BLOCK_LOW_AND_ABOVE = "BLOCK_LOW_AND_ABOVE"
+    BLOCK_MEDIUM_AND_ABOVE = "BLOCK_MEDIUM_AND_ABOVE"
+    BLOCK_ONLY_HIGH = "BLOCK_ONLY_HIGH"
+
+
+class GeminiSafetySetting(BaseModel):
+    category: GeminiSafetyCategory
+    threshold: GeminiSafetyThreshold
+
+
+class GeminiRole(str, Enum):
+    user = "user"
+    model = "model"
+
+
+class GeminiMimeType(str, Enum):
+    application_pdf = "application/pdf"
+    audio_mpeg = "audio/mpeg"
+    audio_mp3 = "audio/mp3"
+    audio_wav = "audio/wav"
+    image_png = "image/png"
+    image_jpeg = "image/jpeg"
+    image_webp = "image/webp"
+    text_plain = "text/plain"
+    video_mov = "video/mov"
+    video_mpeg = "video/mpeg"
+    video_mp4 = "video/mp4"
+    video_mpg = "video/mpg"
+    video_avi = "video/avi"
+    video_wmv = "video/wmv"
+    video_mpegps = "video/mpegps"
+    video_flv = "video/flv"
+
+
+class GeminiInlineData(BaseModel):
+    data: str | None = Field(
+        None,
+        description="The base64 encoding of the image, PDF, or video to include inline in the prompt. "
+        "When including media inline, you must also specify the media type (mimeType) of the data. Size limit: 20MB",
+    )
+    mimeType: GeminiMimeType | None = Field(None)
+
+
+class GeminiPart(BaseModel):
+    inlineData: GeminiInlineData | None = Field(None)
+    text: str | None = Field(None)
+
+
+class GeminiTextPart(BaseModel):
+    text: str | None = Field(None)
+
+
+class GeminiContent(BaseModel):
+    parts: list[GeminiPart] = Field(...)
+    role: GeminiRole = Field(..., examples=["user"])
+
+
+class GeminiSystemInstructionContent(BaseModel):
+    parts: list[GeminiTextPart] = Field(
+        ...,
+        description="A list of ordered parts that make up a single message. "
+        "Different parts may have different IANA MIME types.",
+    )
+    role: GeminiRole = Field(
+        ...,
+        description="The identity of the entity that creates the message. "
+        "The following values are supported: "
+        "user: This indicates that the message is sent by a real person, typically a user-generated message. "
+        "model: This indicates that the message is generated by the model. "
+        "The model value is used to insert messages from model into the conversation during multi-turn conversations. "
+        "For non-multi-turn conversations, this field can be left blank or unset.",
+    )
+
+
+class GeminiFunctionDeclaration(BaseModel):
+    description: str | None = Field(None)
+    name: str = Field(...)
+    parameters: dict[str, Any] = Field(..., description="JSON schema for the function parameters")
+
+
+class GeminiTool(BaseModel):
+    functionDeclarations: list[GeminiFunctionDeclaration] | None = Field(None)
+
+
+class GeminiOffset(BaseModel):
+    nanos: int | None = Field(None, ge=0, le=999999999)
+    seconds: int | None = Field(None, ge=-315576000000, le=315576000000)
+
+
+class GeminiVideoMetadata(BaseModel):
+    endOffset: GeminiOffset | None = Field(None)
+    startOffset: GeminiOffset | None = Field(None)
+
+
+class GeminiGenerationConfig(BaseModel):
+    maxOutputTokens: int | None = Field(None, ge=16, le=8192)
+    seed: int | None = Field(None)
+    stopSequences: list[str] | None = Field(None)
+    temperature: float | None = Field(1, ge=0.0, le=2.0)
+    topK: int | None = Field(40, ge=1)
+    topP: float | None = Field(0.95, ge=0.0, le=1.0)
 
 
 class GeminiImageConfig(BaseModel):
-    aspectRatio: Optional[str] = None
+    aspectRatio: str | None = Field(None)
+    resolution: str | None = Field(None)
 
 
 class GeminiImageGenerationConfig(GeminiGenerationConfig):
-    responseModalities: Optional[list[str]] = None
-    imageConfig: Optional[GeminiImageConfig] = None
+    responseModalities: list[str] | None = Field(None)
+    imageConfig: GeminiImageConfig | None = Field(None)
 
 
 class GeminiImageGenerateContentRequest(BaseModel):
-    contents: list[GeminiContent]
-    generationConfig: Optional[GeminiImageGenerationConfig] = None
-    safetySettings: Optional[list[GeminiSafetySetting]] = None
-    systemInstruction: Optional[GeminiSystemInstructionContent] = None
-    tools: Optional[list[GeminiTool]] = None
-    videoMetadata: Optional[GeminiVideoMetadata] = None
+    contents: list[GeminiContent] = Field(...)
+    generationConfig: GeminiImageGenerationConfig | None = Field(None)
+    safetySettings: list[GeminiSafetySetting] | None = Field(None)
+    systemInstruction: GeminiSystemInstructionContent | None = Field(None)
+    tools: list[GeminiTool] | None = Field(None)
+    videoMetadata: GeminiVideoMetadata | None = Field(None)
+
+
+class GeminiGenerateContentRequest(BaseModel):
+    contents: list[GeminiContent] = Field(...)
+    generationConfig: GeminiGenerationConfig | None = Field(None)
+    safetySettings: list[GeminiSafetySetting] | None = Field(None)
+    systemInstruction: GeminiSystemInstructionContent | None = Field(None)
+    tools: list[GeminiTool] | None = Field(None)
+    videoMetadata: GeminiVideoMetadata | None = Field(None)
+
+
+class Modality(str, Enum):
+    MODALITY_UNSPECIFIED = "MODALITY_UNSPECIFIED"
+    TEXT = "TEXT"
+    IMAGE = "IMAGE"
+    VIDEO = "VIDEO"
+    AUDIO = "AUDIO"
+    DOCUMENT = "DOCUMENT"
+
+
+class ModalityTokenCount(BaseModel):
+    modality: Modality | None = None
+    tokenCount: int | None = Field(None, description="Number of tokens for the given modality.")
+
+
+class Probability(str, Enum):
+    NEGLIGIBLE = "NEGLIGIBLE"
+    LOW = "LOW"
+    MEDIUM = "MEDIUM"
+    HIGH = "HIGH"
+    UNKNOWN = "UNKNOWN"
+
+
+class GeminiSafetyRating(BaseModel):
+    category: GeminiSafetyCategory | None = None
+    probability: Probability | None = Field(
+        None,
+        description="The probability that the content violates the specified safety category",
+    )
+
+
+class GeminiCitation(BaseModel):
+    authors: list[str] | None = None
+    endIndex: int | None = None
+    license: str | None = None
+    publicationDate: date | None = None
+    startIndex: int | None = None
+    title: str | None = None
+    uri: str | None = None
+
+
+class GeminiCitationMetadata(BaseModel):
+    citations: list[GeminiCitation] | None = None
+
+
+class GeminiCandidate(BaseModel):
+    citationMetadata: GeminiCitationMetadata | None = None
+    content: GeminiContent | None = None
+    finishReason: str | None = None
+    safetyRatings: list[GeminiSafetyRating] | None = None
+
+
+class GeminiPromptFeedback(BaseModel):
+    blockReason: str | None = None
+    blockReasonMessage: str | None = None
+    safetyRatings: list[GeminiSafetyRating] | None = None
+
+
+class GeminiUsageMetadata(BaseModel):
+    cachedContentTokenCount: int | None = Field(
+        None,
+        description="Output only. Number of tokens in the cached part in the input (the cached content).",
+    )
+    candidatesTokenCount: int | None = Field(None, description="Number of tokens in the response(s).")
+    candidatesTokensDetails: list[ModalityTokenCount] | None = Field(
+        None, description="Breakdown of candidate tokens by modality."
+    )
+    promptTokenCount: int | None = Field(
+        None,
+        description="Number of tokens in the request. When cachedContent is set, this is still the total effective prompt size meaning this includes the number of tokens in the cached content.",
+    )
+    promptTokensDetails: list[ModalityTokenCount] | None = Field(
+        None, description="Breakdown of prompt tokens by modality."
+    )
+    thoughtsTokenCount: int | None = Field(None, description="Number of tokens present in thoughts output.")
+    toolUsePromptTokenCount: int | None = Field(None, description="Number of tokens present in tool-use prompt(s).")
+
+
+class GeminiGenerateContentResponse(BaseModel):
+    candidates: list[GeminiCandidate] | None = Field(None)
+    promptFeedback: GeminiPromptFeedback | None = Field(None)
+    usageMetadata: GeminiUsageMetadata | None = Field(None)
diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py
@@ -3,35 +3,32 @@
 See: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference
 """
 
-from __future__ import annotations
-
 import base64
 import json
 import os
 import time
 import uuid
 from enum import Enum
 from io import BytesIO
-from typing import Literal, Optional
+from typing import Literal
 
 import torch
 from typing_extensions import override
 
 import folder_paths
 from comfy_api.latest import IO, ComfyExtension, Input
 from comfy_api.util import VideoCodec, VideoContainer
-from comfy_api_nodes.apis import (
+from comfy_api_nodes.apis.gemini_api import (
     GeminiContent,
     GeminiGenerateContentRequest,
     GeminiGenerateContentResponse,
-    GeminiInlineData,
-    GeminiMimeType,
-    GeminiPart,
-)
-from comfy_api_nodes.apis.gemini_api import (
     GeminiImageConfig,
     GeminiImageGenerateContentRequest,
     GeminiImageGenerationConfig,
+    GeminiInlineData,
+    GeminiMimeType,
+    GeminiPart,
+    GeminiRole,
 )
 from comfy_api_nodes.util import (
     ApiEndpoint,
@@ -57,6 +54,7 @@ class GeminiModel(str, Enum):
     gemini_2_5_flash_preview_04_17 = "gemini-2.5-flash-preview-04-17"
     gemini_2_5_pro = "gemini-2.5-pro"
     gemini_2_5_flash = "gemini-2.5-flash"
+    gemini_3_0_pro = "gemini-3-pro-preview"
 
 
 class GeminiImageModel(str, Enum):
@@ -103,6 +101,16 @@ def get_parts_by_type(response: GeminiGenerateContentResponse, part_type: Litera
     Returns:
         List of response parts matching the requested type.
     """
+    if response.candidates is None:
+        if response.promptFeedback.blockReason:
+            feedback = response.promptFeedback
+            raise ValueError(
+                f"Gemini API blocked the request. Reason: {feedback.blockReason} ({feedback.blockReasonMessage})"
+            )
+        raise NotImplementedError(
+            "Gemini returned no response candidates. "
+            "Please report to ComfyUI repository with the example of workflow to reproduce this."
+        )
     parts = []
     for part in response.candidates[0].content.parts:
         if part_type == "text" and hasattr(part, "text") and part.text:
@@ -272,10 +280,10 @@ async def execute(
         prompt: str,
         model: str,
         seed: int,
-        images: Optional[torch.Tensor] = None,
-        audio: Optional[Input.Audio] = None,
-        video: Optional[Input.Video] = None,
-        files: Optional[list[GeminiPart]] = None,
+        images: torch.Tensor | None = None,
+        audio: Input.Audio | None = None,
+        video: Input.Video | None = None,
+        files: list[GeminiPart] | None = None,
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=False)
 
@@ -300,15 +308,14 @@ async def execute(
             data=GeminiGenerateContentRequest(
                 contents=[
                     GeminiContent(
-                        role="user",
+                        role=GeminiRole.user,
                         parts=parts,
                     )
                 ]
             ),
             response_model=GeminiGenerateContentResponse,
         )
 
-        # Get result output
         output_text = get_text_from_response(response)
         if output_text:
             # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button.
@@ -406,7 +413,7 @@ def create_file_part(cls, file_path: str) -> GeminiPart:
         )
 
     @classmethod
-    def execute(cls, file: str, GEMINI_INPUT_FILES: Optional[list[GeminiPart]] = None) -> IO.NodeOutput:
+    def execute(cls, file: str, GEMINI_INPUT_FILES: list[GeminiPart] | None = None) -> IO.NodeOutput:
         """Loads and formats input files for Gemini API."""
         if GEMINI_INPUT_FILES is None:
             GEMINI_INPUT_FILES = []
@@ -421,7 +428,7 @@ class GeminiImage(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="GeminiImageNode",
-            display_name="Google Gemini Image",
+            display_name="Nano Banana (Google Gemini Image)",
             category="api node/image/Gemini",
             description="Edit images synchronously via Google API.",
             inputs=[
@@ -488,8 +495,8 @@ async def execute(
         prompt: str,
         model: str,
         seed: int,
-        images: Optional[torch.Tensor] = None,
-        files: Optional[list[GeminiPart]] = None,
+        images: torch.Tensor | None = None,
+        files: list[GeminiPart] | None = None,
         aspect_ratio: str = "auto",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
@@ -510,7 +517,7 @@ async def execute(
             endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"),
             data=GeminiImageGenerateContentRequest(
                 contents=[
-                    GeminiContent(role="user", parts=parts),
+                    GeminiContent(role=GeminiRole.user, parts=parts),
                 ],
                 generationConfig=GeminiImageGenerationConfig(
                     responseModalities=["TEXT", "IMAGE"],
