Best AI image and video generation skills for AI agents

Ask an agent to "generate an image" or "make a short video" and, working from general knowledge, it will improvise the parts it can't see. It guesses a model id that shifted last quarter, invents an endpoint, reaches for the wrong environment variable, and hands you a curl command that 401s. Worse for video: generation is asynchronous, so an agent that doesn't know to submit-then-poll either hangs, or cheerfully reports success on a job that never rendered. And nothing warns it that a 4K run or a per-second Sora clip costs real money.

The skills here fix that by encoding what the model can't guess. The surprising part is how little the name tells you. A skill called "nano-banana-pro" is not one skill — it is many different documents from many different publishers, and they don't even agree on which Google model they call or which API key they read. The good ones go past "call the API": they bake in the draft-to-final loop, the polling, the cost, and the failure fixes. Those are the ones worth handing your agent.

Top picks

The most disciplined single-model image path

steipete/agent-scripts/nano-banana-pro wraps one script (generate_image.py, run via uv from the user's working directory) and surrounds it with the operational scaffolding most of its namesakes skip. It maps vague requests to real parameters ("hi-res", "ultra" to 4K; nothing said to 1K), pushes a draft-at-1K then final-at-4K loop so you don't burn time rendering 4K on a wrong prompt, and ships a preflight: check uv exists, check the GEMINI_API_KEY is set, check the input file is readable before editing.

What sells it over the near-identical clones is the failure handling — it matches specific API error strings ("No API key provided", quota/403) to concrete fixes, and hands the model generation and editing prompt templates. Caveat: it names the model Gemini 3.1 Flash Image ("Nano Banana 2") and, unlike the openclaw-family variants, exposes no aspect-ratio or multi-image composition flags.

Best for cost tracking and production assets

kingbootoshi/nano-banana-2-skill/nano-banana installs an actual nano-banana CLI (git clone, bun link, a documented /init flow) rather than shelling a bundled script, and it reaches unusually far into real asset work. Every generation logs its spend to ~/.nano-banana/costs.json, viewable with --costs, and the body publishes per-image estimates — roughly $0.067 for a 1K Flash image, about $0.134 on Pro — so the agent knows what it's spending.

Two workflows stand out: a -t flag that generates on a green screen and keys it out with FFmpeg colorkey + despill for clean transparent sprites, and a blank-reference trick for exact pixel dimensions. Caveat: heavier setup (Bun, FFmpeg, ImageMagick), and it was last touched 2026-04-02, older than the current Pro wrappers.

Best for video: VEO and Sora in one script

cdeistopened/skill-stack/video-generator is the most thorough video document I read here. One generate_video.py drives both Google VEO 3.1 and OpenAI Sora behind a --provider flag, and the skill is mostly decision support: use VEO for native synchronized audio, negative prompts, and 4K; use Sora for 12-second clips and visual fidelity. It states the async reality plainly — a job takes 11 seconds to 6 minutes, and the script polls automatically and saves when ready — which is exactly the loop an agent botches on its own.

The prompting guidance is specific: a 150-to-300-character sweet spot, one shot equals one action, colon syntax for dialogue, dialogue lines under 7 words to hold lip-sync, plus a copy-verbatim consistency block for multi-clip work. Caveat: full flexibility wants both API keys, and at 22 stars you're trusting the body, not a crowd.

If you'd rather not marry one model

prime-skills/runcomfy-agent-skills/ai-image-generation takes the opposite bet from the single-model wrappers. It routes across a catalog — FLUX 2, Nano Banana 2 and Pro, GPT Image 2, ByteDance Seedream, Alibaba Qwen and Z-Image, Wan — through the RunComfy runcomfy CLI, one auth and one command, and picks the model by intent: typography precision, photoreal portraits, sub-second drafts, or multi-reference brand styling, each with its own documented prompt patterns.

That flexibility is also the caveat: this is a commercial CLI with an account and credits behind it and utm-tagged links in the doc, so you're adopting a platform, not just a readable script. Its sibling prime-skills/runcomfy-agent-skills/ai-avatar-video extends the same router to audio-driven talking-head and lip-sync models (OmniHuman, Wan via audio_url).

skill	publisher	license	verdict	updated
nano-banana-pro	steipete	MIT	Gemini image wrapper; preflight + draft-to-final + templates	2026-07-23
nano-banana	kingbootoshi	MIT	installed CLI; cost log + green-screen transparency	2026-04-02
video-generator	cdeistopened	MIT	VEO 3.1 + Sora, --provider, auto-polls	2026-07-08
ai-image-generation	prime-skills	MIT	RunComfy CLI router across many image models	2026-05-15
nano-banana-pro	nicepkg	MIT	Gemini 3 Pro wrapper; documents Search-grounded infographics	2026-01-20
nano-banana-pro	evanmopsps	MIT	prompt drafter + affiliate link; runs no local script	2026-06-20

Which nano-banana-pro skill should you install?

Any of them, and none of them — because the name doesn't identify the skill. These are community-published and names aren't namespaced, so "nano-banana-pro" is a topic, not an address. The publisher and the body are the actual identity, and they diverge more than you'd expect for skills sharing a name.

Open the body and read three things. First, the model: steipete's calls Gemini 3.1 Flash Image, while nicepkg/ai-workflow/nano-banana-pro names gemini-3-pro-image-preview outright and the openclaw-flavored beita6969/ScienceClaw/nano-banana-pro calls Gemini 3 Pro Image — a different model, different behavior. Second, the key: most read GEMINI_API_KEY, but a goclaw-flavored variant prefers GOOGLE_API_KEY — its harness metadata requires it — even though its own script still falls back to GEMINI_API_KEY. Whether the variable you hold matters is itself a body-level question. Third, the capabilities: only some expose aspect ratios and multi-image composition (beita6969 composites up to 14 images and prints a MEDIA: line for auto-attach), and nicepkg goes further still, documenting Google Search grounding for data-accurate infographics and declaring its allowed-tools.

Then there's the trap the name sets. evanmopsps/nano-banana-pro-skill/nano-banana-pro carries the same name but runs no local script at all — its own rules tell the agent to return "polished prompts and a Nano Banana Pro link" carrying ref=factory-skill, an affiliate funnel to a hosted site. It's honest about being a prompt drafter, but if you install it expecting the Gemini wrapper the name implies, you get an ad, not a generator. Reading the body is what tells the two apart.

What to check before you install

The same body-first habit protects you across the whole topic, not just the Nano Banana clones.

• Model and key: match the environment variable and model the doc actually names (GEMINI_API_KEY versus GOOGLE_API_KEY, Flash versus Pro) to the key you hold.
• Script or storefront: confirm the skill wraps a script you can read, not a pointer to a hosted platform or an affiliate link.
• Async and cost, for video especially: the trustworthy ones submit and poll for you and state the price — per-second for Sora, per-4K-image for stills — so a long render doesn't hang and a big job doesn't bill silently.
• License: the most common license here is MIT, but the set also holds NOASSERTION and unlicensed entries plus a couple of copyleft AGPL video skills, which matters if generated assets or derivatives ship commercially.
• Setup weight: a drop-in uv script, a git-clone-plus-Bun-plus-FFmpeg CLI, and an account-backed commercial router are three very different commitments.

What to install

The failure you're avoiding is the agent that invents an endpoint, hangs on a video that never rendered, or bills you for a 4K run it never mentioned. For images on a single model, steipete's nano-banana-pro gives the most disciplined path — preflight, draft-to-final, real error fixes. Want spend tracking and transparent game assets, take kingbootoshi's CLI; want to defer the model choice, take the RunComfy router. For video, cdeistopened's generator teaches VEO versus Sora and bakes in the polling loop. Whichever you pick, read the publisher and the body first — the name will not do it for you.