Generate Image — AI Image Creation

You are helping the user generate custom images for their Remotion video project.

Workflow

1. Understand the Need

Ask the user (if not already clear):

• What is the image for? — Scene background, thumbnail, character, product shot, abstract art, diagram?
• Style — Photorealistic, illustration, 3D render, flat design, watercolor, pixel art?
• Dimensions — Landscape (1920x1080), portrait (1080x1920), square (1080x1080)?
• Mood/Colors — Bright, dark, warm, cool, neon, pastel?

2. Craft the Prompt

Write the prompt in full sentences, following Replicate's official prompt-images skill for technique (it's model-agnostic, so it applies to KIE too). Cover, in order: subject, style, lighting, camera angle, color palette, background details.

Good prompt examples:

• "A modern tech startup office seen from above, clean minimalist design with white desks and blue accents, bright natural light from large windows, photorealistic."
• "An abstract geometric pattern of glowing neon purple and blue lines on a dark background, futuristic 3D render."
• "A friendly robot mascot waving, cartoon illustration with pastel colors on a clean white background, suitable for a logo."
• "A downtown skyline silhouetted against a warm orange and pink golden-hour sky, drone aerial shot, cinematic."

For a video, also:

• Leave negative space where Remotion will place titles or captions, and ask for "no text, no watermark"; Remotion renders the text.
• Keep scenes consistent: reuse one style block verbatim across every image in the video, and pass the first good image as a reference. See skills/remotion-production/rules/image-generation.md.

3. Generate via remotion-media (default)

Use remotion-media generate_image:
- prompt: [crafted prompt]
- output_name: img-hero
- aspect_ratio: 16:9  # always pass it (default is 1:1); 9:16 for shorts
- resolution: 2K      # default 1K; use 2K+ for 1080p compositions
- image_urls: [optional reference images, up to 8, public URLs only]

The image is saved to public/img-hero.png in the directory your agent was launched from, so start your agent (Claude Code, Codex, Gemini CLI…) from the Remotion project root. Reference images in image_urls must be public URLs KIE can fetch, not files in public/.

4. Generate via Replicate (specific models)

Use Replicate when the user wants a specific model (typography, several references, a particular look) or KIE fails. It needs the local replicate-mcp server with REPLICATE_API_TOKEN. Follow Replicate's official find-models and run-models skills:

1. Search, then read the schema. search_models with body: "<task>", then get_models (model_owner, model_name) and list_models_examples. Use only inputs that exist in the schema; the aspect-ratio input name varies per model.
2. Create the prediction. Official models: create_models_predictions with model_owner, model_name and input. Community models: create_predictions with version: "owner/name:<version_id>". Fast image models can take Prefer: "wait" (blocks up to 60 s).
3. Poll get_predictions with prediction_id until status is succeeded, failed or canceled.
4. Download the URL in output right away (it expires after 1 hour):

mkdir -p public/images
curl -L -o public/images/img-[descriptive-name].png "<output URL>"

Starting points, as of Oct 2026 (search and read the schema before calling):

• FLUX.2 Pro / Max (black-forest-labs/flux-2-pro, -max): photoreal, several reference images via input_images
• Nano Banana Pro (google/nano-banana-pro): image with text, editing
• GPT Image 2 (openai/gpt-image-2)
• Ideogram v3 Quality / Turbo (owner ideogram-ai, one ID per tier; search ideogram): typography, logos
• FLUX Kontext Pro (black-forest-labs/flux-kontext-pro): edit from a single reference (input_image)

See skills/remotion-production/rules/replicate-models.md for the full flow and rules/image-generation.md for provider choice and consistency guidance.

5. Show Usage in Remotion

Add the image with <CanvasImage> from remotion, as the official remotion-markup skill does (<Img> still works; both accept effects). Sizing and animation follow its images.md and timing.md. The path depends on the provider: generate_image saves to public/<output_name>.png, Replicate downloads to public/images/.

import { CanvasImage, Easing, interpolate, staticFile, useCurrentFrame, useVideoConfig } from "remotion";

const frame = useCurrentFrame();
const { fps } = useVideoConfig();

// As a background (remotion-media output)
<CanvasImage
  src={staticFile("img-hero.png")}
  style={{ width: "100%", height: "100%", objectFit: "cover" }}
/>

// As an element that pops in (Replicate download)
<CanvasImage
  src={staticFile("images/img-product.png")}
  style={{
    scale: interpolate(frame, [0, 0.6 * fps], [0, 1], {
      extrapolateLeft: "clamp", extrapolateRight: "clamp",
      easing: Easing.spring({ damping: 200 }),
      output: "perceptual-scale",
    }),
  }}
/>

For a slow Ken Burns move on a still, see skills/remotion-production/rules/image-generation.md.

6. Iterate

If the image isn't right:

• Adjust the prompt — be more specific about what's wrong (see prompt-images for editing and common pitfalls)
• Generate multiple variations and let the user pick
• Use exclusions only if the model supports them ("no text, no watermark"); some models get worse with negative prompts
• Change one thing at a time so the image still matches the other scenes