Anycap Media Production

AnyCap Media Production

Read this entire file before starting. It covers the full production workflow across image, video, music, and audio -- including iterative refinement with human feedback.

Workflow guide for producing media assets with AnyCap. Covers image, video, music, and audio -- from initial generation through iterative refinement to delivery.

This skill is about how to produce media. For CLI command reference and parameters, read the anycap-cli skill.

Prerequisites

AnyCap CLI must be installed and authenticated. Read the anycap-cli skill if setup is needed.

Quick Reference

All generation commands follow the same pattern:

1. Discover models    anycap {cap} models
2. Check schema       anycap {cap} models <model> schema [--mode <mode>]
3. Generate           anycap {cap} generate --model <model> --prompt "..." -o output.ext

Always choose model IDs from the live model catalog and inspect the schema for the selected mode before relying on model-specific parameters. Always use -o with a descriptive filename.

Image Production

Text-to-Image

Generate an image from a text prompt:

anycap image generate \
  --prompt "a cozy home office with a wooden desk, laptop, coffee cup, and plants by the window" \
  --model <model-id> \
  -o workspace-v1.png

Image-to-Image (Edit / Transform)

Use --mode image-to-image with a reference image to edit or transform an existing image:

anycap image generate \
  --prompt "make it a watercolor painting" \
  --model <model-id> \
  --mode image-to-image \
  --param images=./photo.png \
  -o photo-watercolor.png

Reference images can be local paths or URLs. The CLI handles upload automatically.

Multiple Reference Images

Some models accept multiple reference images for style transfer, composition blending, or subject-driven generation. Use JSON array syntax to pass multiple files:

# Combine style from one image with composition from another
anycap image generate \
  --prompt "merge the architectural style of the first image with the color palette of the second" \
  --model <model-id> \
  --mode image-to-image \
  --param images='["./style-ref.png","./color-ref.png"]' \
  -o blended.png

# Mix local files and URLs
anycap image generate \
  --prompt "a portrait in the style of the reference images" \
  --model <model-id> \
  --mode image-to-image \
  --param images='["./local-ref.png","https://example.com/style-ref.jpg"]' \
  -o portrait-styled.png

Tips:

• Use JSON array syntax '["path1","path2"]' -- repeating --param images= overwrites rather than appends.
• Local file paths inside the array are auto-uploaded, same as single-file mode.
• Not all models support multiple references. Check the model schema first. When unsupported, the model typically uses only the first image.

Iterative Refinement with Annotation

When text prompts alone cannot describe the desired edit precisely ("move this", "remove that specific thing", "change the color of this area"), use the annotation workflow. For the full annotation guide -- including URL/video review, headless access, recording analysis, and multi-user collaboration -- read the anycap-human-interaction skill.

graph TD
    A[Start: concept or existing image] --> B{Have an image?}
    B -->|No| C[Generate initial image]
    B -->|Yes| D[Human annotates the image]
    C --> D
    D --> E[Build prompt from annotations]
    E --> F[Generate with image-to-image]
    F --> G[Show result to human]
    G --> H{Satisfied?}
    H -->|Yes| I[Done -- deliver final asset]
    H -->|No| D

Step 1: Generate or Use an Existing Image

anycap image generate \
  --prompt "a landing page hero banner with mountains and sunrise" \
  --model <model-id> \
  -o banner-v1.png

Step 2: Annotate

Open the annotation tool so the human can visually mark regions, describe desired changes, and optionally record a narrated walkthrough. Multiple users can collaborate on the same session in real-time.

For agent workflows (non-blocking, recommended):

anycap annotate banner-v1.png --no-wait -o banner-v1-annotated.png
# Returns: {session, url, poll_command, stop_command}

Show the URL to the human and ask them to annotate. Multiple people can open the same URL to collaborate. Wait for the human to confirm they are done, then:

# Fetch the result (single call, no loop)
anycap annotate poll --session <session_id>

# If recording exists, analyze it for visual understanding
anycap actions video-read --file .anycap/annotate/<session_id>/recording.webm \
  --instruction "Describe what changes the user wants"

# Clean up
anycap annotate stop --session <session_id>

For interactive sessions (human is at the terminal):

anycap annotate banner-v1.png -o banner-v1-annotated.png
# Blocks until Done click, outputs annotation JSON

The annotation tool supports four tools: Rectangle (R), Arrow (A), Point (P), Freehand (F). Each annotation gets a numbered marker and a text label.

Step 3: Build a Prompt from Annotations

The annotation output contains structured data. Translate each label into a coherent prompt:

{
  "annotations": [
    {"id": 1, "type": "rect", "label": "Replace with a standing desk"},
    {"id": 2, "type": "point", "label": "Add a cat sitting here"},
    {"id": 3, "type": "freehand", "label": "This area should be a bookshelf"}
  ]
}

Prompt: "#1: Replace the desk with a standing desk. #2: Add a cat sitting at the marked position. #3: Transform the outlined area into a bookshelf. Keep all other elements unchanged."

Rules:

• Reference each annotation by its number (#1, #2, etc.)
• Include the human's exact label text
• Add "Keep all other elements unchanged" to preserve unmodified areas

Step 4: Apply the Edit

Use the annotated image (with visual markers) as the reference:

anycap image generate \
  --prompt "#1: Replace the desk with a standing desk. #2: Add a cat. Keep all other elements unchanged." \
  --model <model-id> \
  --mode image-to-image \
  --param images=./banner-v1-annotated.png \
  -o banner-v2.png

Step 5: Iterate

If the human wants more changes, use the latest version as input and repeat from Step 2. Version filenames (v1, v2, v3) so the human can compare and revert.

Image Tips

• Start broad, refine narrow. First generation nails the composition. Annotation iterations handle targeted adjustments.
• One thing at a time. If multi-region edits produce poor results, try one annotation per pass.
• Annotated image only. Pass only the annotated image as the reference. Most models understand numbered markers and remove them from the output.

Video Production

Text-to-Video

anycap video generate \
  --prompt "a cat walking on the beach at sunset, cinematic, slow motion" \
  --model <model-id> \
  -o cat-beach.mp4

Image-to-Video

Animate a still image:

anycap video generate \
  --prompt "gentle camera pan across the landscape, wind blowing through trees" \
  --model <model-id> \
  --mode image-to-video \
  --param images=./landscape.png \
  -o landscape-animated.mp4

This is powerful for combining with image generation: generate a still image first, then animate it.

Video Production Workflow

graph LR
    A[Text prompt] --> B[Generate image]
    B --> C{Animate?}
    C -->|Yes| D[image-to-video]
    C -->|No| E[Done]
    A --> F[text-to-video]
    F --> E