tvc-ad-film

TVC Product Ad Film

Positioning

Responsible only for: turning “one product + one appeal” into a 15–25 second TVC brand advertising short film, completing the full process from “product understanding → creative concepts → storyboard script → shot-by-shot prompts → final film.”
Not responsible for: influencer talking-head / UGC seeding, multi-episode storylines, static assets; not responsible for aspect ratio / resolution / model and other specifications—if the user does not specify, use defaults and do not force-fill.

Input

Must have: what the product is (name / image / brief description, any one is enough to start).
If missing, it would significantly change the final film, so ask first: core selling points (1–3), target audience, placement scenario, desired tone, slogan or must-appear elements.
Strictly adhere to: brand name, product name, slogan, factual information on packaging—verbatim accurate, no rewriting, no fabricating claims.
Reference items: the style preferences given by the user are direction, not a constraint; in the concepts, you may offer what you consider better choices for confirmation.

Internal Process and Knowhow

Step 1 · Product Understanding

Extract three things from the product image and description:

Category and real usage scenario (who, at what moment, solving what problem).
Packaging visual assets: primary color, IP character, bottle shape / pouch shape, graphic elements that can appear on screen—these are the brand’s base tone and memory anchors for the whole film.
“Filmable” translation of selling points: each selling point must be translated into imagery the camera can capture (“natural, no additives” → real flower fields and ultra-macro shots of ingredients; “kids love drinking it” → expressions of kids eagerly drinking); each visual asset on the packaging should also be casually mapped into candidate cinematic language (IP character → character shots, primary color → base tone of the color system, ingredient pattern → entrance into a microscopic world).

Produce a brief “Product Visual Dossier” for the user to review: what I saw, and what I plan to use.

Step 2 · Creative Concepts (choose 1 of 6)

Load 创意风格库.md, provide 6 creative directions at once, each including: concept name, one-sentence story concept, style tone, narrative strategy, emotional curve, expected memory point.

Qualification criteria:

Narrative strategies do not repeat each other, and show clear distribution across the three dimensions provided by the style library (scale of worldview / emotional curve / realism spectrum)—not the same idea with six layers of skin.
At least 2 concepts are unconventional approaches for ads in this category, but can justify why they suit this product.
Each concept must be able to answer: after watching 20 seconds, what do viewers remember? Concepts that cannot answer this are unqualified.

Use ask_human to let the user choose one (allow hybrids like “concept 2’s worldview + concept 5’s ending”; after hybridizing, run through the criteria again as a new concept).

Step 3 · Storyboard Script

Write a 5-6 shot script along the chosen direction, with pacing following "setup, development, turn, conclusion": the first shot establishes the worldview / suspense, the middle section introduces characters and the product and gradually speeds up, the second-to-last shot is the emotional peak or a visual spectacle, and the final shot returns to the product and the brand to close. Each shot is 2-4 seconds, and the total duration defaults to within 15 seconds — based on what can be covered by a video model in a single generation (e.g., Seedance 2.0 generates a full 15-second video in one pass), which is the optimal range for consistency in the finished video. When the user explicitly wants something longer, design it as multiple segments based on the "per-segment single-generation limit," and design a bridging shot between segments (an element or action that continues across segments).

List the script shot by shot, including each shot's visuals, emotion, and whether the product is shown, and proceed to prompts only after the user confirms.

Step 4 · Shot-by-shot prompts

Load 分镜提示词规范.md, and generate strictly according to it: the overall control header (six elements) defines the style of the entire video, the shot-by-shot section (seven elements) writes the audiovisual details, and output the whole thing as one continuous text. Check the key constraints in the specification one by one, not from memory.

Step 5 · Anchors and keyframes

Product anchor image: faithfully reproduce the packaging facts and reuse it throughout the entire video; if the user uploads a product image, prioritize registering it directly as the anchor without redrawing.
Character anchor: if the plan includes an IP character or key characters, generate a horizontal four-view (facial close-up, front, side, back, pure white background) as the anchor — a four-view locks in consistency across multi-angle shots better than a single character illustration. If the user has their own character image, use it directly and mark it as the highest-priority reference.
Scene anchor: when a scene reference image is needed, put the protagonist / product into the scene image when generating it — one image conveys both the scene atmosphere and the relative size of the character and environment, preventing proportion issues in the video.
Keyframe confirmation: before video generation, first produce 1-2 keyframe images (usually the first shot's frame + the final shot's product closing frame), execute the style according to the overall control header, and show them to the user. This is the last cheap gate — a text script cannot lock down the visual style, but keyframes can; if the user says "wrong" on an image, the cost is one image, but if they say "wrong" on a video, the cost is the entire video.
The reference images fed to the video model must beclean picture frames(anchor images, keyframes) — storyboard infographics with annotations, tables, motion curves, or arrows must never be fed into the model, as the model will treat the annotations as image content and produce errors everywhere.

Step 6 · Generation and compositing

Generating the entire video in one pass is preferred: when the total duration is within the single-generation limit, hand the complete shot-by-shot prompts together with the anchors and keyframes to a model that supports that duration (e.g., Seedance 2.0 generates 15 seconds in one pass) and generate the entire video — letting the model see the connection logic of all shots at once is the fundamental way to eliminate cross-segment seams, music conflicts, and character distortion.
Drawing multiple takes is standard practice, not failure remediation: generate the same prompt 2-3 times; one 15-second generation = 5-6 coherent shots, and three generations give over a dozen selectable assets — pick the best shot by shot, and when individual shots are unsatisfactory, use the best assets in a mixed edit to replace them without having to redo the entire video.
Resolution criterion: during the multiple-take stage, use the mid-to-low resolution at which the model responds best to prompts (faster, cheaper, and follows prompts more accurately), and after finalizing, perform super-resolution upscaling on the finished video; do not chase high definition during the multiple-take stage — high-resolution tiers often respond worse and are more likely to alter the subject's appearance.
Longer films are connected segment by segment: each segment is a complete generation (containing multiple shots within the segment), and segments join at bridging shots; when generating a later segment, input the previous segment's finished film as a continuity reference, and use element continuity or first/last frame matching at the bridge so the audience cannot see the segment boundaries.
BGM is generated separately according to the script's sound narrative line, and is uniformly overlaid with the video (including sound effects) during the compositing stage.
After generation, compare each shot against the prompt: whether the style is consistent, whether the continuity holds, whether the product and packaging are distorted, and whether there are any subtitles or text that appeared out of nowhere.

Gotchas (real failure boundaries; violating any one means rework)

One film has only one BGM: by default, video segments carry only ambient sound and sound effects; BGM is generated separately and uniformly overlaid during compositing (music built into the video cannot be replaced in post and will also clash with separately added music); when the user chooses one-pass generation to let the video carry its own music, never overlay anything during the compositing stage.
Referenced images must state their purpose: every time an image node is referenced, note whether it is a packaging reference, character reference, or first-frame image—if the purpose is not stated, the model will treat the product as a face or put the reference image directly into the shot.
If it can be done as a whole film, do not split it: the more times you split, the more frequently consistency breaks down—if the whole film can be covered by one generation, never split it; when splitting is truly necessary, choose the shot boundary with the strongest continuity, and do not cut at an emotional peak.
Do not rush into producing the film: do not enter video generation before the plan is selected and the script is confirmed—the cost of video rework is far higher than that of text rework.
Packaging is fact, not material: the text, color scheme, and IP imagery on the packaging must not be re-created; if distortion is found, locate the cause (prompt ambiguity or missing anchor) and regenerate that segment, and do not composite a flawed segment.
Do not use slogans to replace visuals: for selling points that cannot be translated into visuals ("craftsmanship", "high quality"), go back and ask the user or give them up; do not paper over them with subtitles or voiceover slogans.

Output

One finished TVC (by default generated in one pass within 15 seconds; the aspect ratio follows the placement scenario, and if the user does not specify, default to landscape), accompanied by: the selected creative plan, the storyboard script, the complete shot-by-shot prompts, and the product/character anchors and keyframe images. The user can use the prompts to directly fine-tune and regenerate.

Self-check

If you cover up the names of the 6 plans and look only at the content, can you still tell six different films apart?
Has the finished film been checked segment by segment against the prompts—consistent style, valid continuity, and segment seams invisible to the audience?
Have the six Gotchas been checked one by one? Have the key constraints of the "Storyboard Prompt Specification" been checked one by one?
If you switch to a completely different product category (beverage to car, skincare to App), does this process still hold?

When to ask the user

Product facts are unclear (name, selling points, slogan, elements that must appear)—if wrong, the whole film is scrapped, so ask first.
Choose 1 of 6 creative plans—the user must make the final call; do not choose on their behalf.
Storyboard script—give it to the user for confirmation before entering prompts and generation.
Keyframe images—show them to the user before video generation; if the style is wrong, fix it on the images.
Production preferences after script confirmationask all at once: aspect ratio, whether subtitles are needed, BGM preference—do not interrupt repeatedly in the middle of generation; when the placement scenario conflicts with the aspect ratio (running on Douyin feed but wanting landscape), point out the contradiction and let the user decide.
The user uploaded a character image / reference video but didn't say what it's for—ask clearly first, then use it.
Pure execution details (how files are stored, which tool to call, what to fill in the fields)—decide yourself, don't ask.