Skip to main content

Image Generation Workflows

Build text-to-image pipelines, multi-model comparisons, and image evaluation workflows with Jetty.

Overview​

Image generation is Jetty's most popular use case. This guide covers:

  1. Basic text-to-image generation
  2. Multi-model comparison
  3. Image evaluation with LLM judges
  4. Prompt engineering pipelines
  5. Image modification and video generation

Basic Text-to-Image​

Workflow: Simple Image Generation​

{
  "init_params": {
    "prompt": "A serene lake with mountains at sunset"
  },
  "step_configs": {
    "generate": {
      "model": "black-forest-labs/flux-schnell",
      "activity": "replicate_text2image",
      "prompt_path": "init_params.prompt",
      "output_format": "jpg"
    }
  },
  "steps": ["generate"]
}

Available Models​

ModelQualitySpeedBest For
black-forest-labs/flux-schnellGoodFastQuick iterations
black-forest-labs/flux-kontext-proExcellentMediumProduction images
ideogram-ai/ideogram-v2-turboExcellentMediumText in images
google/imagen-4-fastExcellentFastHigh quality
luma/photon-flashGoodFastArtistic styles

Configuration Options​

{
  "generate": {
    "model": "black-forest-labs/flux-schnell",
    "activity": "replicate_text2image",
    "prompt_path": "init_params.prompt",
    "aspect_ratio": "16:9",
    "output_format": "jpg",
    "num_outputs": 4,
    "guidance_scale": 3.5,
    "num_inference_steps": 24
  }
}

Multi-Model Comparison​

Compare outputs from different models on the same prompt.

Workflow: Side-by-Side Comparison​

Based on production workflow: t2i-evals

{
  "init_params": {
    "prompt": "A beautiful landscape with mountains and a lake"
  },
  "step_configs": {
    "flux": {
      "model": "black-forest-labs/flux-schnell",
      "activity": "replicate_text2image",
      "prompt_path": "init_params.prompt"
    },
    "midjourney": {
      "model": "tstramer/midjourney-diffusion:436b051ebd8f68d23e83d22de5e198e0995357afef113768c20f0b6fcef23c8b",
      "activity": "replicate_text2image",
      "prompt_path": "init_params.prompt",
      "guidance_scale": 7.5,
      "num_inference_steps": 50
    }
  },
  "steps": ["flux", "midjourney"]
}

Batch Model Comparison​

Based on production workflow: t2i-prompt-inversion

{
  "init_params": {
    "prompt": "A charming cottage in a snow blizzard",
    "aspect_ratio": "1:1"
  },
  "step_configs": {
    "ideogram": {
      "model": "ideogram-ai/ideogram-v2-turbo",
      "activity": "replicate_text2image",
      "prompt_path": "init_params.prompt",
      "output_format": "jpg",
      "aspect_ratio_path": "init_params.aspect_ratio"
    },
    "imagen4": {
      "model": "google/imagen-4-fast",
      "activity": "replicate_text2image",
      "prompt_path": "init_params.prompt",
      "output_format": "jpg",
      "aspect_ratio_path": "init_params.aspect_ratio"
    },
    "luma": {
      "model": "luma/photon-flash",
      "activity": "replicate_text2image",
      "prompt_path": "init_params.prompt",
      "output_format": "jpg",
      "aspect_ratio_path": "init_params.aspect_ratio"
    }
  },
  "steps": ["ideogram", "imagen4", "luma"]
}

Image Evaluation with LLM Judges​

Automatically evaluate generated images for quality, compliance, and content.

Workflow: Multi-Criteria Evaluation​

Based on production workflow: llm-judge-plus

{
  "init_params": {
    "prompt": "A wise owl wearing round glasses",
    "model": "black-forest-labs/flux-schnell"
  },
  "step_configs": {
    "generate": {
      "activity": "replicate_text2image",
      "model_path": "init_params.model",
      "prompt_path": "init_params.prompt",
      "aspect_ratio": "16:9",
      "output_format": "jpg"
    },
    "ip_risk": {
      "model": "gpt-4o",
      "activity": "simple_judge",
      "items_path": "generate.outputs.images[0].path",
      "judge_type": "scale",
      "instruction": "There is the potential for an IP Infringement Issue in this image.",
      "scale_range": [0, 1],
      "model_provider": "openai"
    },
    "brand_compliance": {
      "model": "gpt-4o",
      "activity": "simple_judge",
      "items_path": "generate.outputs.images[0].path",
      "judge_type": "scale",
      "instruction": "Evaluate if this image complies with brand guidelines. Check for: proper logo usage, brand color palette adherence, typography consistency, visual style alignment.",
      "scale_range": [0, 1],
      "model_provider": "openai"
    },
    "sensitive_content": {
      "model": "gpt-4o",
      "activity": "simple_judge",
      "items_path": "generate.outputs.images[0].path",
      "judge_type": "scale",
      "instruction": "This image contains sensitive, inappropriate, or potentially harmful content.",
      "scale_range": [0, 1],
      "model_provider": "openai"
    }
  },
  "steps": ["generate", "ip_risk", "brand_compliance", "sensitive_content"]
}

Understanding simple_judge Outputs​

{
  "outputs": {
    "rating": "0.2",
    "explanation": "The image appears to be original artwork...",
    "average_score": 0.2,
    "model": "gpt-4o"
  }
}

Prompt Engineering Pipelines​

Enhance user prompts before image generation.

Workflow: Prompt Expansion​

Based on production workflow: emd-test-1

{
  "init_params": {
    "text": "A scientist doing science",
    "style": "CGI"
  },
  "step_configs": {
    "expand_prompt": {
      "model": "anthropic/claude-3.5-haiku",
      "activity": "replicate_text_stream",
      "max_tokens": 250,
      "prompt_path": "init_params.text",
      "system_prompt": "Take simple inputs and transform them into detailed, visually engaging prompts. Add lighting, colors, and surroundings."
    },
    "generate": {
      "model": "black-forest-labs/flux-schnell",
      "activity": "replicate_text2image",
      "prompt_path": "expand_prompt.outputs.text",
      "output_format": "jpg"
    }
  },
  "steps": ["expand_prompt", "generate"]
}

Combining Multiple Prompt Components​

Based on production workflow: emd-test-1

{
  "init_params": {
    "subject": "A scientist",
    "color_guidance": "Use white background, green and yellow accents",
    "style": "CGI"
  },
  "step_configs": {
    "expand": {
      "model": "anthropic/claude-3.5-haiku",
      "activity": "replicate_text_stream",
      "prompt_path": "init_params.subject",
      "system_prompt": "Elaborate with professional, detailed descriptions."
    },
    "combine": {
      "activity": "text_concatenate",
      "text_paths": [
        "expand.outputs.text",
        "init_params.color_guidance"
      ]
    },
    "generate": {
      "activity": "replicate_text2image",
      "prompt_path": "combine.outputs.text"
    }
  },
  "steps": ["expand", "combine", "generate"]
}

Image Modification​

Modify existing images with AI.

Workflow: Style Transfer​

Based on production workflow: modify_image

{
  "init_params": {
    "prompt": "Make this a 90s cartoon style",
    "image_url": "https://example.com/image.jpg"
  },
  "step_configs": {
    "download": {
      "activity": "download_image",
      "image_url_path": "init_params.image_url"
    },
    "modify": {
      "model": "black-forest-labs/flux-kontext-pro",
      "activity": "replicate_modify_image",
      "prompt_path": "init_params.prompt",
      "input_image_path": "download.outputs.images[0].path",
      "aspect_ratio": "match_input_image",
      "output_format": "jpg"
    }
  },
  "steps": ["download", "modify"]
}

Image to Video​

Generate videos from images.

Workflow: Text-to-Image-to-Video​

Based on production workflow: text2image2video

{
  "init_params": {
    "prompt": "A scientist holding up a balloon",
    "motion": "a zoom out to space"
  },
  "step_configs": {
    "expand": {
      "model": "openai/gpt-4",
      "activity": "litellm_chat",
      "messages_path": "init_params.prompt",
      "system_prompt": "Elaborate the prompt with cyberpunk themes."
    },
    "generate_image": {
      "model": "black-forest-labs/flux-schnell",
      "activity": "replicate_text2image",
      "prompt_path": "expand.outputs.text",
      "output_format": "jpg"
    },
    "animate": {
      "activity": "replicate_text2video",
      "duration": 3,
      "image_path": "generate_image.outputs.images[0].path",
      "resolution": "480p",
      "prompt_path": "init_params.motion"
    }
  },
  "steps": ["expand", "generate_image", "animate"]
}

Batch Image Generation​

Generate images for multiple prompts in parallel.

Workflow: Batch Processing​

Based on production workflow: image-tester

{
  "init_params": {
    "prompts": [
      "Time flies when you're having fun",
      "A portrait of Paris in Paris",
      "A white cat outlined by falling black ink"
    ]
  },
  "step_configs": {
    "generate_all": {
      "activity": "list_emit_await",
      "items_path": "init_params.prompts",
      "task_reference": {
        "task_name": "t2i-evals"
      },
      "data_mapping": {
        "prompt": "{{ $item }}"
      }
    },
    "collect": {
      "activity": "extract_from_trajectories",
      "trajectory_list_path": "generate_all.outputs.trajectory_references",
      "extract_keys": {
        "prompt": "t2i_flux.inputs.prompt",
        "flux_image": "t2i_flux.outputs.images[0].path",
        "midjourney_image": "t2i_midjourney.outputs.images[0].path"
      }
    }
  },
  "steps": ["generate_all", "collect"]
}

Best Practices​

1. Choose the Right Model​

  • Speed: Use flux-schnell for quick iterations
  • Quality: Use flux-kontext-pro or imagen-4 for production
  • Text rendering: Use ideogram for images with text

2. Use Evaluation​

  • Always add simple_judge steps for production workflows
  • Check for IP risk, brand compliance, and sensitive content

3. Optimize Prompts​

  • Use prompt expansion for better results
  • Combine style guidance with subject descriptions

4. Handle Batches Efficiently​

  • Use list_emit_await for parallel processing
  • Collect results with extract_from_trajectories

Next Steps​