カオスマップに戻る
Stable Diffusion

Stable Diffusion

Overview

Stability AI is the creator of Stable Diffusion and the pioneer of open-source image generation, which it has led as a community-driven ecosystem since 2022. The current generation is Stable Diffusion 4 (SD4), announced April 6, 2026. It replaces the U-Net backbone with a diffusion transformer (DiT) and ships in two tiers, SD4 Base and SD4 Ultra.

Current models: Stable Diffusion 4

Pick between Base and Ultra depending on your use case and hardware. Both ship open weights under the community license.

ModelNative resolutionVRAMNotes
SD4 Base1024×102412GB+Streamlined DiT; roughly 8–12s on an RTX 4090. The community workhorse
SD4 Ultra4096×409624GB+Full DiT plus a dedicated text-glyph module and PBR lighting. Enterprise licensing available

Key features

1. Rebuilt on a diffusion transformer

SD4 drops the U-Net backbone for a DiT that mixes local and global attention layers with RoPE spatial embeddings. Structural accuracy — hands in particular — improves sharply over what open models could previously manage.

2. Text inside images (SD4 Ultra)

Ultra includes a dedicated text-glyph module, so logos, signage, and packaging copy render legibly. This was the widest gap between open weights and closed products like Midjourney.

3. Native 4K generation

Ultra generates at 4096×4096 natively (roughly 45–90s on an A100 at max resolution), so print-grade output no longer needs a separate upscaler pass.

4. Deep customization

LoRA and fine-tuning let you build style-specific models, and the community UIs (ComfyUI, Automatic1111) wrap them into production workflows.

5. Stable Audio 2.5 (enterprise audio)

Beyond images, Stability ships Stable Audio 2.5, its first audio model designed for enterprise use. It generates three-minute tracks in seconds on a single GPU and is distributed to WPP clients through WPP Open.

6. EA partnership for game development

A strategic partnership with Electronic Arts covers generated PBR materials and 2D textures plus 3D environment pre-visualization for game production.

License

Stability AI Community License:

  • ✅ Commercial use allowed
  • ✅ Non-commercial use allowed
  • ✅ Free fine-tuning and customization
  • ⚠️ Companies above $1M annual revenue need an Enterprise license

Typical uses

  • Product imagery: catalog shots and ad creative
  • Game assets: characters, environments, UI art, PBR materials
  • Custom models: LoRA / fine-tuning for a specific style
  • Local execution: privacy-sensitive work with no API cost
  • Creative workflows: ComfyUI, Automatic1111, and similar tools

Access and pricing

Running locally is free once you download the weights; the hosted API bills in credits at $0.01 each.

ChannelPrice
Model downloadFree (Community License)
Stability AI API — Stable Image Core$0.03/image (3 credits)
Stability AI API — Stable Image Ultra$0.08/image (8 credits)
API membership$20/mo for 6,000 credits
Amazon BedrockAvailable through AWS

Strengths

  • Pioneer of open-source image generation
  • Commercial use permitted (Community License)
  • SD4 Base runs on a consumer GPU (12GB VRAM)
  • Highly customizable (LoRA, fine-tuning)
  • SD4 Ultra does native 4K and legible in-image text
  • Active community (Hugging Face, Civitai)

Weaknesses

  • Business instability (rebuilding after a leadership change)
  • SD4 Ultra needs 24GB+ VRAM, which is a stretch for most personal setups
  • Limited video models compared with Runway or Kling
  • Weak text models — the focus is image generation

Official links

公式サイト