Establish anchors once. For each recurring entity — each character (@Name), each key location (#Name), the overall style — generate ONE locked reference: a clean full-body/portrait on a plain background with even lighting for a character; an empty establishing plate for a location. Write one exhaustive canonical description (age, body, face, skin, hair, every garment + color, 1–2 distinguishing features) and keep it — you will quote its key traits verbatim wherever the entity appears. Cap ~4 cast and ~4 settings. A storyboard's story bible also yields these (see flowy-storyboard).
Pass the anchor into every dependent generation via input_media (generate_image/generate_video/create_node/update_node take up to 9 refs). This — not the canvas edge — is what actually conditions a generation. On a video, set each image's role so it plays the part you intend:
- role: "reference" — identity/style/continuity (characters, products, the look). Required for 3+ images. Use this for anchors.
- role: "first_frame" / "last_frame" — the clip literally starts/ends on that still (boundary frames for continuous motion).
- Omit role and the server guesses by count: 1–2 images become first/last frame, 3+ become references — so a single anchor with no role is treated as a start frame, not a style reference. Pass role: "reference" to anchor a look. reference and first/last can't be mixed in one call. On a canvas, wire anchors as EDGES — the graph-native form of this rule: create_node(connect_to: [{node_id: <anchor>, role: "reference"}, …]) draws the edge AND feeds the anchor's output into the generation in one call; on an existing node, connect_nodes then update_node (a regeneration consumes the outputs of all incoming edges). Anchors that only ride input_media condition the result but leave the canvas graph unwired — prefer connect_to inside a project.
Prompt only the delta. Describe action, framing, and lighting — not the identity (that comes from the reference). Keep appearance text next to a reference MINIMAL: long re-description fights the image and degrades it. When you do name the subject, use the SAME words every time — synonym drift ("a young woman" → "the stylish creator") makes the model re-interpret the identity.
Word-lock the small stuff anyway. Logos, printed text, insignia, exact brand colors get dropped from references — restate each as a positive lock at the end of the prompt ("the jacket patch reads 'APEX' in orange stitching"), even though it's visible on the anchor.
Reference only where present. An anchor passed into a shot where that subject doesn't appear gets forced into frame.
Never re-roll an anchored subject from a bare text prompt — that mints a new, inconsistent subject. Iterate with update_node (which carries the node's current image + refs unless fresh: true) or regenerate with the same input_media.

Chained beat images (motion continuity)

For a scene broken into per-second beats, generate the beat stills as a CHAIN: beat k's image is conditioned on beat k−1's output (pass it first in input_media, then the character/setting anchors). Consecutive frames then differ only by what one beat of time produces — pose, expression, framing — while identity, wardrobe, palette, and lighting carry over. Feed the whole chain to the video generation as references.

Style continuity

Hold lighting, palette, lens language, and grade across a scene by reusing the SAME look words in every shot prompt (one short look line, not a big style prefix — long verbatim prefixes dilute attention and drift the output generic). Keep a style reference image in input_media when the look must be exact.

Guidance

Signed asset URLs expire — re-fetch with get_asset (or re-read the node via get_node) before reusing an anchor in a later session.
Verify continuity by LOOKING at each result next to its anchor (see flowy-verify); if the face/wardrobe drifted, the fix is refs (wire the anchor, strengthen its precedence), not more prose.

flowy-continuity

Flowy: continuity anchors

Method