피파 한 줄 정리: Open (FLUX·SD·Voxtral): 영구성·프라이버시·custom. Closed (MJ·GPT-Image·Runway): 품질·convenience. Sora 사례 = closed 의존도의 위험.
Think of open models as owning a kitchen vs. closed models as ordering from a restaurant. The restaurant (closed model) handles everything — ingredients, prep, cooking, plating — and you get a polished result. But you can't control the recipe, the ingredients might change without notice, and the restaurant might close (ask Sora users). Your own kitchen (open model) requires more work, but you control everything, run it on your schedule, and nobody can take it away.
The Open-Source Tier
FLUX.1 Schnell (Apache 2.0, commercial use allowed) is the most accessible high-quality open model. At 12B parameters with 4-step generation, it runs on consumer GPUs and produces results competitive with closed APIs. FLUX.2 Dev at 32B parameters pushes quality higher but requires more compute. FLUX.2 Klein variants (4B–9B) enable real-time generation on modest hardware (~13GB VRAM).
Stable Diffusion offers the deepest ecosystem: thousands of community fine-tunes, LoRA adapters, ControlNet extensions, and workflow tools like ComfyUI. The base model trails FLUX in raw quality, but the customization ecosystem is unmatched.
For video, LTX-2 and Wan 2.6 provide open alternatives at 4K and 1080p respectively. For voice, Voxtral runs on a single 16GB GPU.
The Closed/Proprietary Tier
Midjourney, GPT-Image, Imagen 4, Runway Gen-4.5, Veo 3.1, and ElevenLabs are all API-only or platform-only services. They typically offer higher baseline quality, simpler interfaces, and less setup. But you're subject to their pricing, content policies, availability, and continuity decisions.
The Decision Matrix
- Open models give you permanence, privacy, and customization. Closed models give you quality, convenience, and low setup cost.
- Sora's shutdown proves that even billion-dollar-backed closed models can disappear overnight.
- A hybrid approach — closed for exploration, open for production — often offers the best of both worlds.