Flow Image Gen

flow-image-gen

Generate all storyboard images for one short-form video job using Google's Gemini image model. Self-contained — one curl per image, no external services beyond the Gemini API.

Inputs

A job folder path, e.g. examples/demo-job/. Inside it, input.json with:

• image_prompts[] — array with id, prompt, optional negative_prompt
• timeline[] — used to map each prompt id to its output filename (.image)
• resolution — like "1080x1920" — used to derive aspect ratio
• image_style — object with consistency_anchor and color_grade (both appended to every prompt) for visual consistency across the set

image_style is read from input.json and merged into each prompt automatically.

Quick run

The bundled runner implements the whole loop (parallelism, retries, skip-existing, PNG size check). Run it as one invocation:

JOB=examples/demo-job bash skills/flow-image-gen/scripts/gen_images.sh

Tune concurrency with IMAGE_GEN_PARALLEL (default 4). The steps below document exactly what that script does, in case you want to drive it by hand.

Stage gate (status.json)

This skill reads and writes /status.json to support idempotent re-runs.

Gate check — run BEFORE any other step

STATUS_FILE="$JOB/status.json"

# Initialize if absent (one-time, first skill to run)
if [ ! -f "$STATUS_FILE" ]; then
  jq -n '{
    schema_version: 1,
    stages: {images: "pending", voiceover: "pending", render: "pending"},
    artifacts: {images_completed: 0, voiceover_duration_ms: null, output_path: null},
    errors: []
  }' > "$STATUS_FILE"
fi

STAGE_STATUS=$(jq -r '.stages.images // "pending"' "$STATUS_FILE")
if [ "$STAGE_STATUS" = "done" ]; then
  echo "Skipped (images stage already done)"
  exit 0
elif [ "$STAGE_STATUS" = "failed" ]; then
  echo "FAILED: images stage previously failed. Check status.json. Exiting." >&2
  exit 1
fi

# Mark running
jq '.stages.images = "running"' "$STATUS_FILE" > "${STATUS_FILE}.tmp" && mv "${STATUS_FILE}.tmp" "$STATUS_FILE"

On success — write done + artifact count

IMAGES_COUNT=$(ls -1 "$JOB/images/"*.png 2>/dev/null | wc -l)
jq --argjson count "$IMAGES_COUNT" \
  '.stages.images = "done" | .artifacts.images_completed = $count' \
  "$STATUS_FILE" > "${STATUS_FILE}.tmp" && mv "${STATUS_FILE}.tmp" "$STATUS_FILE"

On any failure — write failed + error

jq --arg msg "" \
  '.stages.images = "failed" | .errors += [{"stage": "images", "message": $msg, "time": (now | strftime("%Y-%m-%dT%H:%M:%SZ"))}]' \
  "$STATUS_FILE" > "${STATUS_FILE}.tmp" && mv "${STATUS_FILE}.tmp" "$STATUS_FILE"

Provider details

• Endpoint: https://generativelanguage.googleapis.com/v1beta/models/gemini-3.1-flash-image-preview:generateContent
• Auth: header x-goog-api-key: $GEMINI_API_KEY (NOT Authorization: Bearer — Google rejects that with ACCESS_TOKEN_TYPE_UNSUPPORTED)
• Response shape: base64-encoded PNG inside candidates[0].content.parts[].inlineData.data
• Cost: ~$0.067 per image at 1K resolution

Steps (what the runner does)

1. Confirm GEMINI_API_KEY is set. If empty, exit with error.

2. Read /input.json. Parse image_prompts[], timeline[], resolution, image_style.

3. Derive aspect ratio from resolution:

   • "1080x1920" → "9:16" (vertical Shorts)
   • "1920x1080" → "16:9" (landscape)
   • "1080x1080" → "1:1" (square)

4. Ensure /images/ exists.

5. Extract image_style.consistency_anchor and image_style.color_grade. These two strings are appended to every prompt for visual consistency.

6. Build the work list: for each image_prompts[i], the destination filename comes from the matching timeline[].image (matched by id, else positional, else .png).

7. Generate the images — up to 4 in parallel, never sequentially one curl at a time (a 14-image job drops from ~4 min to ~1 min of waiting). Per image:

   a. Build the destination path: $JOB/images/.

   b. Skip if it already exists and is non-empty (supports re-runs after partial failure).

   c. Build the request body, merging the style anchor + color grade into the prompt:

      REQ_FILE=$(mktemp)
      FULL_PROMPT="${PROMPT}. ${STYLE_ANCHOR} Color grade: ${COLOR_GRADE}."
      jq -n --arg p "$FULL_PROMPT" --arg ar "$ASPECT_RATIO" '{
        contents: [{parts: [{text: $p}]}],
        generationConfig: {responseModalities: ["IMAGE"], imageConfig: {aspectRatio: $ar}}
      }' > "$REQ_FILE"

d. Submit the request:

      RESP_FILE=$(mktemp)
      HTTP_CODE=$(curl -sS -X POST \
        "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.1-flash-image-preview:generateContent" \
        -H "x-goog-api-key: $GEMINI_API_KEY" \
        -H "Content-Type: application/json" \
        -d @"$REQ_FILE" -w "%{http_code}" -o "$RESP_FILE")

e. Error check. If HTTP code is not 200, OR the response JSON has an .error key, report and retry once, then fail loudly.

f. Decode the base64 image and write to disk:

      jq -r '.candidates[0].content.parts[] | select(.inlineData) | .inlineData.data' "$RESP_FILE" \
        | base64 -d > "$DEST"

g. Verify the file is a valid PNG and non-trivially sized (>= 10 KB).

h. Clean up temp files; print OK $DEST.

1. After the loop, print Generated N/M images. and exit non-zero if any image failed.

Notes

• The responseModalities: ["IMAGE"] field is required. Without it, the same model returns a text description of the image instead of the image bytes.
• negative_prompt from image_prompts[] and image_style.negative_prompt are not used as a native negative parameter (this model doesn't support one). If exclusion matters, append them to the main prompt as natural language: "no text overlay, no watermark."
• Aspect ratio enum values the model accepts: "1:1", "3:4", "4:3", "9:16", "16:9". Anything else 400s.
• Each image costs roughly $0.067 at 1K resolution. A 14-image Short costs ~$0.94 in API calls.
• If you see ACCESS_TOKEN_TYPE_UNSUPPORTED, the auth header is wrong (must be x-goog-api-key, not Authorization: Bearer).
• If you see RESOURCE_EXHAUSTED with FreeTier quotas, billing isn't active on the project that owns the key.

Output

Per image: OK to stdout. Final summary line: Generated N/M images. On failure: FAILED (id=): to stderr, exit non-zero.