agentleFS
vargHQvargHQ/skills/product-photo-to-video/SKILL.md

product-photo-to-video

Skill34 starsChanged 8 months agoMITView on GitHub

Turn a product photo into a product video ad that keeps the real product. Builds TikTok, Reels, TikTok Shop b-roll and listing videos from one to three product photos: a product truth sheet (label text letter by letter, colours, parts), a cut-out, a hero still, an 8–12 panel storyboard sheet, then one multi-shot Seedance 2.5 or Kling 4 clip with a PRODUCT LOCK block, a repair pass on drifting shots, and the real packshot composited on the last second. Formats: hands-only ASMR unboxing, macro pour or texture, ingredient burst or exploded view, product in use, 360° loop, product in a hand, apparel try-on. Never ships a Ken Burns slideshow. Use when someone has product photos and wants a product video, packshot animation, unboxing, ASMR product video, listing video or AI b-roll without filming. NOT for a person talking about the product (use ai-ugc-ad-video), TikTok Shop presenter videos (use tiktok-shop-video), or still listing images (use a product-photo skill).

  • Reads credentials

The file

We split this skill into the rules it gives an agent

1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Check at 100%, letter by letter, colour by eyedropper.

    line 63

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Turn a product photo into a product video ad

What wins in late 2026 is one impossible but tactile moment built around the real product: a macro pour, a lid snap, an ingredient burst, a hands-only unboxing. What fails: logo-first "cinematic" b-roll, still-photo slideshows (TikTok Shop bans them in shoppable videos), and any frame where the model rewrote the label.

Generation runs on varg (MCP call_tool or POST https://api.varg.ai/v2/tools/{image|video}/call). Quote with estimate_cost and wait for a yes before each paid step. A still costs cents to fix; a clip costs dollars, so approve stills first.

1. Product truth sheet (free, do it first)

From the photos and the product page, write:

PRODUCT: <name, size, e.g. "30 ml amber glass dropper bottle, gold cap">
LABEL (exact, in order): "NORTHLINE" / "VITAMIN C" / "30 ML"
COLOURS: body #F4A23B, cap #C9A227, label #FFFFFF
PARTS / IN THE BOX: bottle, dropper, box
SCALE: fits in the palm
CLAIMS WE CAN SHOW: absorbs in ~30 s (visible), lightweight texture
NEVER: change shape, size, colour or material; add text; fake results (fizz, instant clearing)

Photos: clean, straight-on, label readable. "If the label isn't readable in the reference, the model invents its own text."

2. Cut-out and hero still

  1. Cut-out: bria_guided_remove_background (keeps the original pixels).
  2. Hero still: gpt_image_2_5 edit, "keep the uploaded product exactly unchanged, replace only the background with [environment], match reflections and contact shadows, no product deformation, no added labels". nano_banana_2_1 is cheaper but may need small packaging text fixed.
  3. 9:16 canvas: bria_fibo_edit_1_5_expand. Real side views for reveals: qwen_angles (exact angles, not invented sides).
  4. Check at 100%, letter by letter, colour by eyedropper. Reject and redo, don't "fix later".

Product in a hand: bria_fibo_edit_1_5_product_holding. Apparel on a person: bria_fibo_edit_1_5_virtual_try_on (check stitching, print and logo; AI try-on smooths texture and length).

3. Storyboard sheet (approve before video)

gpt_image_2_5, 8–12 panels, each ≥1.5 s with one camera move, POV hands only for no-face formats. Template: references/prompts.md P1. Sequence that works: product on the surface, hands enter → open and reveal → the most tactile moments → in use → the result in macro → hero packshot. Show it to the person.

4. One multi-shot clip with PRODUCT LOCK

seedance_2_5, 10–15 s, @image1 = the cut-out, @image2 = the approved sheet ("follow its shot order, never show its panels, numbers or text"), audio on for foley, "no background music" unless wanted, last second held clean for the end card. Template: references/prompts.md P2.

Cheaper: kling_v4 multi-shot (often right first time) or wan_3 (best value for b-roll volume). Drafts: kling_v4_flash.

Safest for listings and loops: pixelcut_looping_video (locked camera, seamless loop, motion: spin for 360°). Exact orbit around a packshot: minimax_h3_max_camera_controls.

5. Repair, don't reroll

Reject any shot where the pack, label, shape, colour, part count or hand count drifts. Regenerate only that shot from first and last frames (seedance_2_5 first/last-frame mode, kling_v4 end_image_url), camera locked in both frames or it invents a move. Reliability ladder, best first:

  1. Composite the real product PNG and label over AI footage (real pixels)
  2. Locked camera, only the environment moves (liquid, fog, hands)
  3. Reference on every shot + PRODUCT LOCK + label spelled out
  4. One job per reference ("defines the product ONLY")
  5. First/last frames from approved stills
  6. A hand anchored on the box base; hands at the product's edges
  7. Push-ins and parallax instead of full spins of unseen sides

Common failures and fixes: references/failures.md.

6. Composite the truth and export

  • Lay the real packshot PNG (and copy, if any) over the clean last second:

    # packshot.png = the real cut-out with alpha; 13 = where the clean last seconds start
    ffmpeg -i clip.mp4 -loop 1 -i packshot.png -filter_complex "[1:v]scale=760:-1[p];[0:v][p]overlay=(W-w)/2:(H-h)/2:shortest=1:enable='gte(t,13)'" -c:v libx264 -crf 18 -c:a copy hero.mp4

    Tested on FFmpeg 6.0: the packshot appears from 13 s and the clip keeps its length.

  • Edit, captions, CTA, loudness and the 4:5 cut: ecommerce-video-edit. A 1:1 or 16:9 15–30 s cut with no hard sell for the product page.
  • Label realistic AI footage (TikTok AIGC toggle; Amazon contains-synthetic-performer if a photoreal person appears). Never show the product doing more than it can.

Formats

FormatModelNotes
Hands-only ASMR unboxingseedance_2_5 / kling_v412-panel ASMR sheet; "natural ASMR only"; one hand stays on the box base
Macro pour / texture / splashseedance_2_5Locked camera on the pack; only the liquid moves
Ingredient burst / exploded viewgpt_image_2_5 still → seedance_2_5Real parts in their real order
Food process film ending on the productseedance_2_58–12 beats of 2–3 s, match cuts, 110–120 BPM
360° loop for listingspixelcut_looping_videoSafest label fidelity
Product in use (kitchen, bathroom)nano_banana_2_1 stills → kling_v4Hands at edges, real scale cue
Apparel on a modelbria_fibo_edit_1_5_virtual_try_on → seedance_2_5One @image per garment, "EXCLUSIVE reference"

Costs on varg (live rules, Oct 2026; quote exact jobs with estimate_cost): a 15 s product video is roughly 250–1,000 credits at 720p (Seedance 2.5 is the priciest and varies ~2× by route; Kling 4 ~16 credits/s, Wan 3 ~10 credits/s at 720p).

Related

ecommerce-video-ads · varg-product-hook (UGC-style hook clip of the product in use) · ecommerce-video-edit · ai-ugc-ad-video · tiktok-shop-video · varg-ai

Sources: varg how-to: turn a product photo into a video, fal commercial skill, Masonry controlled test, Luma Seedance guide, OpenAI image prompting cookbook, X posts by @abxxai, @Ciri_ai, @HeyRu0by, @Strength04_X, @yourPlugAI.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.