create-video-seedance-2-fal

create-video-seedance-2-fal

⚠️ REQUIRED PREFLIGHT (2026-06-04): every Seedance prompt that names a branded product MUST be paired with a real reference image of that exact product as one of the --image-ref inputs. Text-only product description without a ref invites Seedance to invent geometry that does not match the real SKU. If no reference exists locally, find one via brand-assets → existing-ads → PDP harvest → promote to clients/<brand>/brand-assets/reference-photos/. Optionally clean an occluded ref via a quick GPT-image-2 edit pass first. Always get user review of the reference + final prompt before firing the paid call. See feedback_product_reference_required.md.

Purpose

Wraps the FAL endpoint bytedance/seedance-2.0/reference-to-video. Pass 1-N image references (typically: portrait + product + optional setting) + a structured prompt → get back a 4-15s clip with native lip-synced VO and ambient audio.

Seedance 2.0 renders multiple internal cuts within a single 15s call when the prompt specifies sub-scene structure (e.g. WIDE HOOK / PRODUCT HERO / SIDESTEP / REACTION). This is the architectural win over the older NB2 + i2v multi-keyframe approach — fewer calls, native lip-sync, no Soul ID required.

Use this atom when:

• The clip needs a talking-head avatar reviewing or demonstrating a product
• Native lip-synced dialogue is required (set --generate-audio)
• Identity must hold via image reference, not video reference (see Decision Rule 1)
• Length 4-15s, vertical 9:16 (or other aspect ratios)
• Multi-image conditioning (face + product, or face + multiple products for a hook scene)

Do NOT use for:

• > 15s clips → split across multiple calls and stitch (assembly/stitch-videos-ffmpeg)
• Pure cinematic camera moves without an on-screen person → create-video-veo3 is often better
• Product-only B-roll with no person → create-product-videos-higgsfield-ms or text-to-video Seedance
• Clips that need video reference for continuity from a prior AI-gen scene → not supported (content_policy_violation)

Inputs

Required:

• --prompt — structured prompt block (see "Prompt template" below). Long, multi-block. Reference run example: beauty-by-earth/video-01-three-product-grwm/working/fire_seedance_facewash.py.
• --output — local MP4 destination.
• --image-url (alias --image-ref) — at least one PUBLIC reference image URL (repeatable), passed as image_urls. Order matters — first = @Image1 in prompt addressing. The proxy does NOT upload local files: host local refs via MCP get_upload_url → get_download_url and pass the URL (identical to create-video-fal).

Optional:

• --resolution — 480p | 720p | 1080p (default). 1080p meaningfully better for product label fidelity.
• --duration — 4–15 seconds (default 15). Passed as string per FAL schema.
• --aspect-ratio — 9:16 (default), 16:9, 1:1, etc.
• --generate-audio — bool, default true. Set false for silent B-roll where VO is added post.
• --seed — integer for deterministic re-runs (FAL returns a seed; pass it back to reproduce).

Credentials:

• No FAL key. Routes through the GooseWorks FAL proxy (media_proxy.py, bundled) and bills the Ads agent, using ~/.gooseworks/credentials.json (written by the gooseworks CLI). Your cal_/agent token is not a FAL key — the old direct-key path 401'd; that's why this capability was rerouted through the proxy.

Decision Rules

0. Prefer FEWER, LONGER calls with internal multi-scene prompting over many short calls. Every new Seedance call drifts character, wardrobe, lighting, and background slightly — even with the same portrait reference. 5 short calls = 5 drift points the viewer registers as continuity breaks. Use the proven multi-scene template inside ONE 10-15s call to get internal hard-feeling cuts from a single render:

SCENE 1 (0-3s) — WIDE HOOK [framing + action + dialogue]
SCENE 2 (3-7s) — PRODUCT HERO [framing + camera move + dialogue]
SCENE 3 (7-11s) — SIDESTEP / PROOF [framing + action + dialogue]
SCENE 4 (11-15s) — REACTION + CTA [framing + action + dialogue]

Setting CAN vary across internal sub-scenes within one Seedance call. Character identity is locked by the portrait reference image, not by the setting. Validated reference (ref-01 Bristle UGC, 26s) shows the same creator across 5+ different room/background settings within a single multi-scene render. Don't artificially split into multiple calls just to change rooms — bake the setting changes INTO the SCENE N (X-Ys) sub-scene structure of a single call. Distribute calls based on STORY ARC and the 15s max-duration cap, not setting count. See prompt-example.md for the validated template. Memory: feedback_seedance_long_calls_not_short.md.

0a. WARDROBE must be FULLY specified — top + bottom + hair + accessories. Validated 2026-05-23 (Bristle): if only the top is specified ("cream sweatshirt"), Seedance hallucinates the bottom (denim shorts appeared from nowhere). Every prompt must lock all of:

• Top (with negations)
• Bottom (with negations) — "light denim cuffed shorts, mid-thigh, NOT jeans, NOT pajama shorts"
• Hair (specific styling — "loose low ponytail at the nape with face-framing strands" works better than vague "bun")
• Accessories (e.g. "thin gold hoop earrings, nothing else")

For confessional / single-sitting UGC, wardrobe should be IDENTICAL across all calls of the same ad (validated user preference + ref-01 + ad-03 patterns). Vary wardrobe across calls ONLY for GRWM / transformation / multi-day formats. Memory: feedback_seedance_long_calls_not_short.md.

0b. PASS A BACKGROUND REFERENCE IMAGE as image_urls[2] for locked settings. Validated 2026-05-23 (Bristle). Text-only setting description yields generic-rental aesthetic — Seedance fills with its own room prior. A background hero image (generated via NB2 Pro at fal-ai/nano-banana-pro, $0.15/1K, text-to-image with "no bokeh, infinite focus, no people, no faces" directives) LOCKS the room across all sub-scenes. Use the SAME background image across all calls of the same ad — different bg image per call reads as "she moved." Memory: feedback_seedance_long_calls_not_short.md.

0c. PRE-CLEAN PRODUCT REFERENCE IMAGES via NB2 edit if they contain accessories. Validated 2026-05-23 (Bristle hi_res-12 had vials + funnel beside the box; Seedance fused them onto the box face). Inspect source product images at intake. If accessories / secondary products / score badges / lifestyle elements visible, route through fal-ai/nano-banana-2/edit ($0.08) to isolate the hero product before passing to Seedance. Cheaper than rerolling a $4.50-$10.20 Seedance render. Memory: feedback_nb2_clean_product_refs.md.

0d. SEQUENTIAL CALL CONTINUITY — use seed reuse, NOT video_urls. Validated 2026-05-23 (Bristle Variant A/B test). For Call N>1 in a multi-call ad, pass the prior call's seed (via --seed) with the same image_urls. NEVER pass the prior call's mp4 as video_urls for sequential continuity — it HURTS continuity (model tries to "continue" prior video; wardrobe/product drift more than they otherwise would). video_urls is ONLY for style transfer from a REAL HUMAN reference video (e.g. take a real-creator UGC ad and recreate with AI creator). Memory: feedback_seedance_video_urls_vs_seed.md.

0e. WHEN USING video_urls (style transfer from real reference): set client timeout to 30+ min AND trim reference to ≤15s. Validated 2026-05-23:

• Render time: ~18-20 min (vs ~10 min for image-only). The default fal_client.subscribe() 10-min timeout will fail. Use fal_client.submit() + manual polling.
• Combined video_urls duration cap: 15s. FAL rejects with input_value_error: Combined video duration must not exceed 15.0s. Trim reference videos with ffmpeg -t 15 before upload.
• Combined files cap: 50 MB total. Our fal_helpers.upload_file() cap bumped from 10 MB to 50 MB on 2026-05-23.
• Detect validation errors (input_value_error / "must not exceed") in polling exceptions and bail after 1-2 retries — terminal errors don't surface via status=Failed.
• Video reference transfers MORE than style/pacing — it transfers wardrobe styling, accessories, hair to a non-trivial degree (validated 2026-05-23 Probiotic Recreation: hair flipped to match reference; tablets appeared despite cleaned product image). Text-side overrides are PARTIALLY effective; pick a reference whose visuals already align with your target.
• FAL DOES accept AI-gen video as input (the old "instant content_policy_violation" assumption from BBE 2026-05 early is outdated). Memory: feedback_seedance_video_urls_vs_seed.md.

1. NEVER pass AI-generated video as video_urls. FAL's content-policy validator rejects this with partner_validation_failed. The 60% video-input discount is real but only applies to REAL HUMAN reference video. For AI-avatar continuity across multiple scenes, use the same identity portrait as image_urls in every scene. Validated on BBE.

2. NSFW reject → STOP and surface, do not auto-retry. Body-application + female + water/lather hits the classifier reliably. If FAL returns content_policy_violation or status: failed with nsfw reason, do NOT retry the same prompt. Surface to the caller with 3 options: rewrite the application beat as smell-test / fingertips-show, reframe as POV (no face), or skip the scene. Follows the project-wide moderation policy (see memory feedback_hf_moderation_surface.md).

3. duration as string. Pass "15", not 15. FAL schema accepts string only.

4. Single product per call (except hook). Multi-product hero scenes within a single Seedance call degrade label fidelity. The exception is the hook/intro scene where all products appear together but no single label is hero.

5. 1080p for hero quality. Skin texture and product label crispness scale noticeably with resolution. Use 720p only for budget/probe calls.

Workflow

python3 scripts/generate.py \
  --prompt "$(cat prompt-scene-1.txt)" \
  --output /path/to/scene-1.mp4 \
  --image-url "https://<hosted>/portrait.png" \
  --image-url "https://<hosted>/product.png" \
  --resolution 1080p \
  --duration 15 \
  --aspect-ratio 9:16 \
  --generate-audio

The script:

1. Routes through the bundled media_proxy.py (GooseWorks FAL proxy) — no FAL key; bills the Ads agent via ~/.gooseworks/credentials.json.
2. Passes the --image-url values straight through as image_urls (they must already be PUBLIC URLs; the orchestrator hosts local refs via MCP).
3. Submits bytedance/seedance-2.0/reference-to-video through the proxy and polls to completion with: ```python { "prompt": <prompt>, "image_urls": [<refs>], "resolution": <res>, "duration": "<dur>", # string "aspect_ratio": <ar>, "generate_audio": <bool>, "seed": <optional int>, } ```
4. Downloads the result MP4 to --output.
5. Writes <output>.meta.json with prompt, refs, video URL, seed, duration, cost estimate.
6. On content_policy_violation or NSFW failure: surfaces the error and exits non-zero (no silent retry).

Output

• <output> — MP4 clip at requested duration, resolution, aspect ratio.
• <output>.meta.json — request + result metadata + seed for reproducibility + cost estimate.

Pricing (2026-05)

| Resolution | Duration | Cost per call | |---|---|---| | 720p std | 15s | ~$4.54 | | 1080p std | 15s | ~$10.20 | | 1080p std | 8s | ~$5.44 | | 1080p std | 4s | ~$2.72 |

Fast tier is ~20% cheaper but caps at 720p.

Prompt template

The prompt is the single biggest quality lever. Use the proven BBE template structure:

📱 UGC INFLUENCER [PRODUCT] REVIEW REEL (SELFIE-STYLE, NOT A COMMERCIAL)

[BLOCK 0 — anti-pattern opener]
This is NOT a cinematic commercial, NOT a [category] ad, NOT an editorial campaign.
This is a real Instagram Reel filmed by a 20-something [demographic] creator on her
iPhone front camera, talking directly to her followers in a relaxed, confiding tone...

[IMAGE REF ASSIGNMENTS]
Use @Image1 as the influencer's face/identity. Use @Image2 as the exact [product]
she is reviewing — it is a [VISUAL DESCRIPTION]. It is NOT a [common confusion].
The product and label MUST be clearly identifiable.

[FORMAT block]
- Vertical 9:16 phone-shot selfie aesthetic, slight handheld micro-shake
- WARDROBE (STRICT): [wardrobe]. ABSOLUTELY NOT [list of negations].
- Setting: [setting]
- Lighting: [lighting]
- Real, slightly imperfect skin (natural barely-any makeup), light golden tan

[IDENTITY block]
Same young woman as @Image1 — [age range, hair, eyes, skin tone]

[DYNAMIC FRAMING block]
The video must vary its framing across scenes like a real Reel — do NOT stay
on one focal length the whole time. Mix wide / medium / close-ups.

[SCENE 1 (0-3s)] WIDE HOOK — [script line] + [emotion + action]
[SCENE 2 (3-7s)] PRODUCT HERO with rotation — [script line] + [emotion + action]
[SCENE 3 (7-11s)] SIDESTEP application or smell-test — [script line] + [emotion + action]
[SCENE 4 (11-15s)] REACTION + recommendation — [script line] + [emotion + action]

[AUDIO DIRECTION]
[Voice character: tone, accent, pace per DELIVERY_STYLES.md]
NO music in the audio (we add music in post).

[GLOBAL RULES]
- Wardrobe identical across the scene
- Product stays identical and clearly visible in her hand
- Lips remain CLOSED between dialogue beats
- [NSFW sidestep negations if relevant]

[STRICT REALISM RULES — Block F]
- Tiny natural imperfections, subtle peach fuzz, faint fine lines
- NOT plastic, NOT airbrushed, NOT doll-like, NOT porcelain
- Skin highlights ONLY on nose tip, cheekbones, cupid's bow
- Hair: individual flyaway strands, slight frizz
- Image grain matches iPhone front camera

Full template reference: prompt-example.md at the repo root, plus all four scripts in beauty-by-earth/video-01-three-product-grwm/working/fire_seedance_*.py.

Quality Checks

• Output MP4 exists, plays end-to-end, duration within ±0.5s of --duration.
• Aspect ratio and resolution match the request.
• Avatar identity matches --image-ref portrait reference (visual check).
• Product label readable in the hero sub-scene (zoom in on the t=mid frame).
• Lip-sync within ~100ms of the audio.
• No NSFW visual leaked (no water/lather/body-application unless explicitly approved).
• meta.json records gateway: "fal", model: "bytedance/seedance-2.0/reference-to-video", and the FAL seed.

Failure Modes

| Failure | Cause | Recovery | |---|---|---| | FAL 422 content_policy_violation: partner_validation_failed | AI-gen scene passed as video_urls | Remove video_urls. Use image_urls only for identity continuity. | | FAL response status: failed, nsfw reason | Body-application + female + water/lather hit classifier | STOP. Surface to caller. Rewrite sidestep beat as smell-test or fingertips-show. NEVER auto-retry. | | FAL 404 on endpoint path | Wrong endpoint slug | Use bytedance/seedance-2.0/reference-to-video exactly — no fal-ai/ prefix. | | Output longer than requested | duration sent as int not string | Send as string: "duration": "15". | | Product label garbled / wrong | Multi-image refs drifted, or 720p | Move to 1080p, ensure product ref is sharp + correctly cropped, add "label MUST be sharp and clearly readable" to the hero sub-scene. | | Mouth moving when not speaking | Default Seedance behavior | Add to Global Rules: "Lips remain CLOSED between dialogue beats — no mouth movement when not speaking." | | Cinematic slow pacing despite "iPhone selfie" prompt | Seedance prior leans cinematic | Strengthen anti-pattern block: "Casual real-time pace, NOT slow-motion, NOT dolly moves, NOT cinematic." Optionally post-process with ffmpeg setpts=0.66*PTS for 1.5x speed. | | Heavy freckles / blotchy skin | Block B language too aggressive | Use v8 soft-skin language. Drop "freckles prominent", "real-person imperfections". Replace with "tiny natural imperfections, subtle peach fuzz, faint fine lines". | | Wrapper RuntimeError "fal subscribe exceeded timeout" AFTER long elapse (>600s on 1080p) | Shared fal_helpers.subscribe() post-call timeout check raises even when the result returned successfully (latent bug; default bumped to 2400s on 2026-05-26 but recovery still applies for older runs) | The render likely completed on FAL — DO NOT re-fire. See Recovery Workflow below. | | FAL downstream_service_unavailable after long wait (~30 min) | Seedance backend overloaded / transient infra issue | Not billed (FAL bills on completion). Retry once. If retry also fails, wait 30 min then retry. Don't change prompt — it's infra, not your input. | | Brand name mispronounced in native-audio output (validated 2026-06-08 on Alitu A3: rendered "Alitu" as "al-too" instead of "ah-LEE-too") | Seedance picks its own voice + interprets non-English / made-up brand names phonetically without dictionary knowledge | Preflight: use phonetic spelling directly in the dialogue (NOT a parenthetical). E.g. SPOKEN LINE: "Ali-too does the engineering for you." Avoid Dialogue: "Alitu (pronounced ah-LEE-too)" — Seedance reads the parenthetical out loud (see next row). Post-render QC: /watch:watch and confirm the Whisper transcript matches expected pronunciation. Recovery: re-roll the affected sub-scene with phonetic spelling baked into the line, OR splice an ElevenLabs reading of the brand-name beat over the Seedance visual. | | Prompt metadata leaks into spoken line (validated 2026-06-08 on Alitu A3 v2: rendered "Alitu, pronounced Ali-tu, three syllables, does the engineering for you" — Seedance read the parenthetical voice-coaching out loud as if it were part of the dialogue) | Seedance does not distinguish "spoken text" from "voice coaching" inside a Dialogue: field. Parentheticals, notes, "(pronounced X)", "(softly)", "(stage whisper)" all get verbalized. | Prevention: keep voice-coaching OUTSIDE the dialogue field entirely. Use a separate SPOKEN LINE: heading whose content is the EXACT verbatim utterance, no parens, no notes. Put coaching under NEGATIONS: as "DO NOT say X / DO NOT explain pronunciation". Bake phonetic intent into the dialogue spelling itself (Ali-too not Alitu). |

Recovery Workflow — recover a lost result via FAL share URL

When a Seedance render is paid for but the result was discarded by the wrapper / shell / timeout, the mp4 is still accessible via FAL's dashboard for 24-48+ hours.

1. Ask the user to open https://fal.ai/dashboard/requests (or the share-link view of the model: https://fal.ai/models/bytedance/seedance-2.0/reference-to-video?share=<uuid>) signed in to the FAL account that holds FAL_API_KEY.
2. Filter by model = bytedance/seedance-2.0/reference-to-video and look for the request matching the lost run's timestamp.
3. The user copies the share URL (contains ?share=<uuid>) and pastes it back.
4. WebFetch the share URL extracting "the v3.fal.media/files/.../*.mp4 URL plus the seed and any other metadata". The page renders all of this verbatim.
5. curl -sSL -o <output>.mp4 "<extracted-mp4-url>" to grab the file. Zero extra cost.
6. Save metadata to <output>.mp4.meta.json (seed, share URL, mp4 URL, recovery note) so downstream calls can use the seed for continuity.

Prevention: Fire long Seedance calls with --with-logs so queue-update messages including the request_id stream to the atom's stdout. If the wrapper later errors, the request_id is recoverable from the background task output file at /private/tmp/claude-501/.../tasks/<id>.output and can be passed to fal_client.result(request_id) directly.

Memory: feedback_fal_helpers_timeout_bug.md. Validated: Lineage Video 01 Call 1 (2026-05-26) — recovered $8.83 render via share URL.

References

• FAL model page: https://fal.ai/models/bytedance/seedance-2.0/reference-to-video
• Reference runs: beauty-by-earth/video-01-three-product-grwm/working/fire_seedance_{facewash,deo_probe,hook,tanner}.py
• Proven prompt template: prompt-example.md
• Sibling NB2 atom (avatar / product stills): coworkers/video/atoms/image-generation/create-image-nano-banana-2-fal
• Composing molecule: coworkers/video/molecules/ugc-ad/create-ugc-style-video
• Shared helpers: coworkers/video/atoms/_shared/fal_helpers.py