Loading content...
Create stunning AI art on your own hardware
Understanding the different image generation models helps you choose the right tool for your work.
The classic. 512x512 native resolution. Runs on 6-8GB GPUs. Massive ecosystem of fine-tunes, LoRAs, and embeddings. Best for: anime, specific styles via community models.
Major upgrade. 1024x1024 native. Needs 10-12GB VRAM. Much better composition and prompt following. Optional refiner for extra detail. Best for: general purpose high quality.
Latest from Black Forest Labs (SD creators). Best text rendering. Superior prompt understanding. Needs 16-24GB VRAM. Best for: highest quality, text in images, complex prompts.
Stability AI's answer to Flux. Good text rendering. 12GB minimum. Better than SDXL, different from Flux. Best for: balanced quality/requirements.
Image generation is VRAM-intensive. Here's what different GPUs can handle.
Runs SD 1.5 well. SDXL possible with optimizations (VAE tiling, fp16). Cannot run Flux. ~3-5 images per minute with SD 1.5.
SDXL runs comfortably. Flux Schnell possible. ~5-8 images per minute with SDXL. Good for most users.
All models including Flux Dev. Comfortable batch sizes. ~8-12 images per minute. Recommended for serious work.
Maximum speed and quality. Large batches, no compromises. ~15-25 images per minute. Training LoRAs viable.
Different interfaces serve different needs.
Node-based workflow editor. Most powerful and flexible. Steeper learning curve. Preferred by professionals. Required for advanced techniques.
Traditional web interface. Easier than ComfyUI. Good extension ecosystem. Less flexible for complex workflows.
Simplified Midjourney-like experience. Minimal settings. Good for beginners. Limited customization.
Balance of power and usability. Good for intermediate users. Canvas for inpainting.
Unlock the full potential of local image generation.
Guide image generation with reference images. Pose, depth, edges, and more. Essential for consistent characters and scenes.
Small additive models that modify style or add subjects. Thousands available on CivitAI. Can train your own on 12GB+ GPUs.
Edit specific parts of images. Extend images beyond original boundaries. Essential for iterative refinement.
Increase resolution post-generation. Models like 4x-UltraSharp. Can go from 1024 to 4K+ with detail.
Improve your results with these proven techniques.
Be specific about style, lighting, composition. Use quality boosters: 'masterpiece, best quality, highly detailed'. Negative prompts to exclude unwanted elements.
DPM++ 2M Karras is reliable. 20-30 steps for drafts, 40-50 for finals. CFG 7-8 for balance, lower for creative freedom.
Generate many variations quickly. Use img2img to refine favorites. Inpaint problem areas. Upscale final results.
Generate at lower resolution first. Pick best compositions. Regenerate winners at high resolution. Much faster than high-res from start.
Check our step-by-step setup guides and GPU recommendations.