Imported from jeremylongshore/tons-of-skills-marketplace (skills/.curated/klingai-performance-tuning/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill klingai-performance-tuning. Copyright stays with the author (MIT).

Kling AI Performance Tuning

Overview

Optimize video generation for your use case by choosing the right model, mode, and parameters. Covers benchmarking, speed vs. quality trade-offs, connection pooling, and caching strategies.

Speed vs. Quality Matrix

Benchmarking Tool

import time, requests, json

def benchmark_model(prompt: str, model: str, mode: str = "standard",
                    runs: int = 3) -> dict:
    """Benchmark generation time for a model/mode combination."""
    times = []

    for i in range(runs):
        start = time.monotonic()

        # Submit
        r = requests.post(f"{BASE}/videos/text2video", headers=get_headers(), json={
            "model_name": model, "prompt": prompt, "duration": "5", "mode": mode,
        }).json()
        task_id = r["data"]["task_id"]

        # Poll
        while True:
            time.sleep(10)
            result = requests.get(
                f"{BASE}/videos/text2video/{task_id}", headers=get_headers()
            ).json()
            if result["data"]["task_status"] in ("succeed", "failed"):
                break

        elapsed = time.monotonic() - start
        times.append(elapsed)
        print(f"  Run {i+1}/{runs}: {elapsed:.1f}s ({result['data']['task_status']})")

    return {
        "model": model,
        "mode": mode,
        "avg_sec": round(sum(times) / len(times), 1),
        "min_sec": round(min(times), 1),
        "max_sec": round(max(times), 1),
        "runs": runs,
    }

# Compare models
prompt = "A waterfall in a tropical forest, cinematic"
for model in ["kling-v2-5-turbo", "kling-v2-master", "kling-v2-6"]:
    result = benchmark_model(prompt, model, runs=2)
    print(f"{model}: avg={result['avg_sec']}s, min={result['min_sec']}s")

Connection Pooling

import requests

# Without pooling: new TCP connection per request (slow)
# With pooling: reuse connections (fast)

session = requests.Session()
adapter = requests.adapters.HTTPAdapter(
    pool_connections=5,     # number of connection pools
    pool_maxsize=10,        # max connections per pool
    max_retries=3,          # auto-retry on connection errors
)
session.mount("https://", adapter)

# Use session instead of requests directly
response = session.post(f"{BASE}/videos/text2video", headers=get_headers(), json=body)

Prompt Optimization

Prompts that generate faster:

# Slow prompt (vague, conflicting)
slow = "A scene with many things happening, realistic but also artistic"

# Fast prompt (specific, clear)
fast = "A single red fox walking through snow, side view, natural lighting, 4K"

Caching Strategy

import hashlib

class PromptCache:
    """Cache results to avoid regenerating identical videos."""

    def __init__(self):
        self._cache = {}

    def _key(self, prompt: str, model: str, duration: int, mode: str) -> str:
        raw = f"{prompt}|{model}|{duration}|{mode}"
        return hashlib.sha256(raw.encode()).hexdigest()[:16]

    def get(self, prompt, model, duration, mode):
        key = self._key(prompt, model, duration, mode)
        return self._cache.get(key)

    def set(self, prompt, model, duration, mode, video_url):
        key = self._key(prompt, model, duration, mode)
        self._cache[key] = {
            "url": video_url,
            "cached_at": time.time(),
        }

cache = PromptCache()

def generate_with_cache(prompt, model="kling-v2-master", duration=5, mode="standard"):
    cached = cache.get(prompt, model, duration, mode)
    if cached:
        print(f"Cache hit: {cached['url']}")
        return cached["url"]

    # Generate
    result = client.text_to_video(prompt, model=model, duration=duration, mode=mode)
    url = result["videos"][0]["url"]
    cache.set(prompt, model, duration, mode, url)
    return url

Optimization Checklist

• Use kling-v2-5-turbo for iteration, v2-6 for final
• Use standard mode until final render
• Connection pooling via requests.Session()
• Cache identical prompt+param combinations
• Prompt: specific, single subject, < 200 words
• Batch submissions paced at 2-3s intervals
• Use callback_url instead of polling
• Download videos async (don't block on CDN download)

Prerequisites

• An approved performance baseline, synthetic or rights-cleared test brief, sandbox workspace, content-policy review, credit cap, draft-only destination, and rollback/removal owner.

Instructions

1. Benchmark caching, model selection, batching, and callback changes using bounded watermarked sandbox drafts only.
2. Capture aggregate latency, error, task, and credit metrics; verify policy, rights, destination, retention, and removal controls before comparison.
3. Halt the canary on quality, policy, rights, budget, scope, or retention drift and restore the prior configuration.
4. Promote tuning changes only after owner approval; retain a redacted benchmark receipt and remove temporary assets at the approved boundary.

Output

Produce a performance receipt with environment, baseline and aggregate measurements, model/configuration category, policy/rights/budget checks, draft-only assertion, owner approval, retention/removal proof, and rollback reference. Exclude prompts, assets, identities, and secrets.

Error Handling

Examples

env=staging; brief=synthetic; p95_delta=-18%; credits=within-cap; policy=pass; destination=draft-only; rollback=available supports approval.

Resources

• Model Catalog
• Developer Portal