Local vs Cloud Vision

Where local vision wins

Privacy is paramount. Medical images, legal documents, surveillance footage, employee data — these often cannot go to the cloud. Local is the only option.
Bulk batches. "Process 50,000 product photos" is free locally, expensive on the cloud.
Offline. Field work, air-gapped sites, edge deployments.
Latency-sensitive UI. Local vision can answer in 1–2 seconds; cloud round-trips add 3–5+ seconds.

Where cloud still wins

Hardest reasoning over images. "Read this diagram and explain the architecture" — Claude 3.5 Sonnet and GPT-4V are still ahead.
Very high resolution. Cloud models handle 4K+ images more gracefully.
Spatial reasoning at the frontier. Counting, occlusion, relative position — cloud is better.

The Pippa pattern

cwkPippa uses local vision (Gemma 3 / Qwen 2.5-VL) for routine image tasks (avatar generation prompts, screenshot debugging, OCR of receipts) and falls back to Claude vision for hardest cases. Local-first with cloud-fallback applies to vision identically to text.

Code

A local-first vision router·python

async def describe_image_smart(path: str, complexity: str = "auto") -> str:
    """Try local first; escalate to cloud for hard cases."""
    # Step 1: local
    local_answer = await ollama_vision(path,
                                       "Describe this image in detail.")

    # Step 2: confidence check (heuristic — if the local answer is too short
    # or contains hedge phrases, escalate)
    if complexity == "always_cloud" or _is_low_confidence(local_answer):
        return await claude_vision(path,
                                   "Describe this image in detail.")
    return local_answer

def _is_low_confidence(text: str) -> bool:
    return (len(text) < 80
            or any(h in text.lower() for h in [
                "i can't", "unable to", "not sure", "hard to tell",
            ]))

Where local vision wins

Where cloud still wins

The Pippa pattern

Code

External links

Exercise

Progress

Comments 0