Generator Nodes

Generator nodes create new content - images, video, 3D models, audio, and text. Every model available on Scenario is accessible as a generator node in workflows - all of them. This includes the Gemini family, Seedream family, GPT family, Flux family, Kling family, Seedance family, Veo, Wan, and any custom-trained or uploaded models in your project. For the full, up-to-date list of models, see scenario.com/models.

📷 [Screenshot: The Generator nodes section of the Nodes Panel showing Image Generator, Video Generator, 3D Generator, Audio Generator, and LLM Node]

Image Generator

Creates images from text prompts, reference images, or a combination of both. Supports both generation (creating new images) and editing (modifying existing images).

Inputs

Outputs

Settings

• Model: Every image model on Scenario is available here: the Gemini family (Gemini, Gemini 3.1), the Flux family (Flux, Flux 2 Pro), the Seedream family, GPT Image, Ideogram, and any custom-trained or uploaded models in your project. New models are added regularly.

• Aspect Ratio: Output dimensions (1:1, 16:9, 9:16, 3:2, 2:3, and more)

• Resolution: Pixel dimensions, varies by model (up to 2048x2048)

• Image Count: Number of images to generate per run (1 to 10)

• Seed: Optional. Set a specific seed to reproduce identical results. Leave blank for variation.

• Use Google Search: When enabled, the model can reference web search results to improve prompt understanding.

The Image Generator handles both generation (text-to-image or reference-guided creation) and editing (modifying an existing image based on a text instruction such as "change the background to a beach"). Connect a reference image and provide an editing instruction in the prompt to use editing mode.

Key Models

• Gemini / Gemini 3.1 - Strong at following complex, multi-part prompts

• Seedream - Specialized for photorealistic outputs

• Flux / Flux 2 Pro - Fast, high-quality general-purpose generation

• Ideogram - Excels at text and typography in images

• GPT Image - OpenAI's image generation model

• Custom-trained and uploaded models - Any LoRA or fine-tuned model you have trained or uploaded in your Scenario project is available in the model dropdown

🎬 Video Generator

Creates video from text prompts, start/end frames, reference images, or existing video. Supports text-to-video, image-to-video, and video-to-video workflows. Over 90 video models are available.

Inputs

Outputs

The First Frame and Last Frame outputs extract the opening and closing frames of the generated video as images. These can be fed into downstream nodes. For example, chaining the Last Frame of one video into the First Frame of the next creates continuous sequences.

Settings

• Model: 90+ video models available across all major model families: Kling (Kling O1, Kling V3, Kling O1 Video Editing), Seedance, Veo (from Google), Wan, LTX, Pika, P-Video, and more. New models are added regularly.

• Duration: Video length (typically 4 to 8 seconds, varies by model)

• Aspect Ratio: Output dimensions (1:1, 16:9, 9:16, and more)

• Negative Prompt: Describe what you do not want in the video

• Reference Frames: Connect First Frame and/or Last Frame images to control the start and end points of the generated video

Workflow modes:

• Text-to-video: Provide only a prompt

• Image-to-video: Connect a start image (First Frame) to animate from a specific frame

• Video-to-video: Connect an existing video as input for style transfer or editing (for example, Kling O1 Video Editing)

• Frame-guided: Provide both First Frame and Last Frame to control start and end points

Some models support combining video with additional elements such as audio, images, or overlays. For example, Kling elements allow you to attach reference images or audio that influence the generated video. Check model-specific settings for available element inputs.

🧊 3D Generator

Creates 3D models from reference images, text prompts, or existing 3D assets. Supports generation, editing, texturing, rigging, retopology, and splitting models into parts.

Inputs

Outputs

Settings

• Model: Options include Hunyuan, Meshy, Trellis, Tripo P1, Tripo 3.1, Polygen 1.5, Tencent Texture Edit, Tencent UV Unwrapping, Tripo Retopology, Tripo 3.0 Texturing, and more

• Resolution: Output quality and detail level

• Output Format: FBX or GLB depending on model

The 3D Generator covers multiple operations depending on the selected model:

• Generation: Create a 3D model from a reference image or prompt

• Multi-image view: Generate from multiple reference angles

• Editing: Modify an existing 3D model

• Texturing: Apply or refine PBR textures on a model

• Rigging: Add skeletal rigs for animation

• Retopology: Optimize mesh topology for game engines or real-time use

• UV Unwrapping: Generate UV maps for texture painting

• Splitting in parts: Decompose a model into separate components

Not all models support all operations. Tencent Texture Edit requires FBX input specifically.

Operations Supported

• Generation: Image-to-3D or Text-to-3D using Trellis or Tripo 3.1.

• Post-Processing: Tencent UV Unwrapping, Tripo Retopology (for game-ready meshes), and Rigging.

• Texturing: Apply PBR materials or use Tencent Texture Edit for specific re-texturing tasks.

• Rigging: Generate rigs for bipedal characters using the Meshy Rigging or Tripo Rigging models.

• Unwrapping: Optimize the UV maps of the textures to make editing easier.

🎵 Audio Generator

Creates audio content from text, images, or existing audio. Covers voice generation, music, sound effects, and voice cloning.

Inputs

Outputs

Settings

• Model: Options include ElevenLabs, Beatoven, TTS, Lux TTS, Tada, MMAudio 2, HeyGen Video Translate, and more

• Voice Selection: Available for TTS models

• Duration: For music generation

• Additional parameters vary by model

Generation modes:

• From text: Text-to-speech or text-to-music. Provide a script or description.

• From image: Generate audio that matches or describes an image (using multimodal audio models like MMAudio 2)

• From audio (voice cloning): Provide a reference audio sample to clone a voice or match a style

Modes of Operation

• Text-to-Speech (TTS): Using ElevenLabs or Lux TTS.

• Multimodal (Image-to-Audio): Use MMAudio 2 to generate soundscapes that match the visual content of an image.

• Music Generation: Create full tracks via Beatoven or MiniMax Music.

🧠 LLM Node: The Workflow Brain

Runs a large language model within the workflow to generate, transform, or analyze text and images. Often used for prompt engineering: analyzing reference images and generating detailed prompts that feed into other generators.

Inputs

Outputs

Settings

• Model: Options include the Google Gemini suite (Gemini 2.5 Flash, Gemini 3 Flash Preview), the OpenAI GPT suite (GPT-4.1 Mini, GPT-5 Mini), and other text and vision models. New models are added regularly.

• Number of Results: How many text outputs to generate (1 to 10)

• Thinking Level: Controls reasoning depth, balancing cost, speed, and output quality:

  • Minimal: Fastest. Best for simple formatting or basic text transformations.

  • Medium: Balanced. Ideal for standard prompt generation and style analysis.

  • High: Best reasoning. Necessary for complex world-building, intricate character descriptions, and deep logical branching.

The Instruction field is the core configuration. Write it as a system prompt: define a role, describe the task, and set constraints.

Common patterns:

• Prompt generation: LLM analyzes a reference image, generates a detailed prompt, feeds it into Image Generator or Video Generator

• Content-aware style transfer: LLM compares a style reference and content image, generates a transfer prompt

• Text transformation: LLM rewrites, summarizes, or reformats text

• Image analysis: LLM describes or captions images (requires a vision-capable model)

• Multi-step reasoning: Chain multiple LLM Nodes for iterative refinement

• Style Guard: Set up an LLM as a style guard agent by providing a style reference image and instruction. The LLM analyzes the reference and generates prompts that ensure every image in the workflow stays within the project's visual language.

Thinking Levels

• Minimal: Lightning-fast. Best for formatting or simple rewrites.

• Medium: The standard for prompt engineering and style analysis.

• High: Maximum reasoning. Use this for complex world-building or when you need the node to act as a Style Guard to ensure visual consistency across a 50-node graph.

Example Workflow

The LLM analyzes a "Character Reference" image of a warrior. Based on its Instruction, it generates a highly detailed output: "A female elven battle mage, with long silver hair... wearing agile, form-fitting armor... only the same art style and design as the reference image should be used".

This detailed output is then automatically fed into an Image Generator node, ensuring the final asset perfectly matches your project's aesthetic.

Choosing a Generator