Photograph the stout faucet you sell: Vidu Q3 image-to-video is 5s, 10s, or 15s at 1080p, 16cr discounted, native audio for the cascade only, coffee-oem already spent the machine logo, overlay the restrictor disc as authored type, and /for/coffee-roasters is the roastery not this tap SKU.
Nano Banana 2 is 8cr discounted: lock the nylon pyramid mesh still before any fill motion; paper tea-bag tags already spent Qwen on the tag line, so this leftover is the pyramid geometry.
Flux 3 first-last interpolates two stills you locked, 5–20s at 9 credits with native audio. Pixverse Transition is 5 / 8 / 10s at 1080p, 40 credits, audio unset — a clip-to-clip morph, not two catalog stills. A pack-shot A/B is first-last; a trailer smash is Transition.
Seedream 5.0 Pro at 7cr letters the Pulumi registry provider card — package name and version; GitHub Action marketplace already owns the badge row, so do not Omni a cluster walkthrough.
Reddit's 2026 crackdown is on AI-generated spam in threads, not on hosting an AI video you made. Do not read a comment-spam rule as a ban on Seedance or Kling.
The older Seed Dance I2V page is 1080p, 2–12s, 9 credits, audio null — do not treat it as Seedance 2.5 (47cr, native audio) and do not invent a /models/seedance-2-5-image-to-video URL.
The published Seedance v1.5 Pro T2V row is 1080p, native audio, 4–12s at 6 credits — a different family page from Seedance 2.5 at 47 credits / 720p / 4–30s.
Snapchat said it will no longer recommend wholly AI-generated videos in Spotlight. The generate can still exist; the recommendation slot is what moved.
veo3-1-extend-video is extend-video at 7s, 720p, from 112 credits, native audio. It continues a keeper; it is not a new Veo 3.1 generate (from 64 credits).
Marketplace extension listings already spent the icon-plus-screenshot row; this leftover is the theme token plate, which GPT Image 2 (4K, 4cr discounted) can letter — do not Omni a talking developer, and I2V is optional after the tokens hold.
Wan Video 2.5 I2V and T2V both bill 15 credits at 1080p for 5s or 10s; the image-to-video row has audio false, the text-to-video row has native audio — same family, different meter flag.
wan-v2-6-text-to-image is 3 credits at 4K; the sibling Wan Video 2.5 T2I row is 5 credits at 4K — two still pages, two rates, neither is a video generate.
A seated bung and a pulled bung on the same wine barrel interpolate on Veo first-last at 40cr, 4K, 4s / 6s / 8s, native audio; /for/wineries is the cellar page, not this stave SKU.
The setup surface lists ten hosts that take the same MCP URL and skill pack: Claude Code, Cursor, Codex, Gemini CLI, Windsurf, Copilot, Cline, Continue, Goose, and Replit — the hero only rotates five of them.
Flat-billed image models charge a fixed price per completed call — aspect ratio, resolution, and prompt length are not terms. 100 keepers is 100 calls only if every prompt lands; a 50% keep rate doubles the leftover /cost/100-ai-product-images meter. Per-megapixel rows are a different formula.
The /templates catalog is 15 one-tap files: start_template_run needs hosted input URLs, usually one still; Finger Snap is the cheapest at 22 credits, Red Carpet is the only row with two optional person stills, and none of them take a scene list.
The /workflows catalog is 17 scene graphs you clone with recreate_public_template, saved not auto-run; Morning Matcha Routine is the unique 405/405 row (preview equals full), Lunchwell Hospital Visit is the 1800-credit ceiling.
Of nine caption families (45 presets), only Snap (Anton) and Reveal (TikTok Sans) are word-timed; the other seven land the whole line as a subtitle block — bookmark /caption-styles/group/snap and /caption-styles/group/reveal as one karaoke index, not nine twin posts.
Hailuo 2.3 Fast, Standard, and Pro publish two duration tiers — 6s and 10s — so a 7-second idea buys the 10s SKU. Seedance 1 Pro Fast is the same rounding as 5-second blocks. That leftover staircase is /cost/7-second-clip-on-a-tiered-model, not a per-second trim.
Versely is three objects, not one plugin: paste https://mcp.versely.studio/mcp, run npm install -g @versely/cli then versely install, and npx skills add AI-XLabs-Innovation/versely-skills — skills are procedures, MCP is the callable, the CLI wires both.
Only three of 17 public recipes pin Kling Video V3 Pro Image to Video — Morning Matcha 405/405, NYC Street Interview 504/504, Versely Podcast Promo 270/270 — and they are exactly the three where preview equals full. The other 14 recipes pin a different video model and roughly double on the full meter.
Three of the 17 public recipes estimate under 200 credits: Viral Panel Reaction Reel at 60, Podcast With Me at 156, and Plush Lipstick Influencer Review at 186 — they are still cloned scene graphs, not one-tap start_template_run files.