Hyper-SD

image

by ByteDance · Updated October 1, 2025

Hyper-SD is ByteDance's novel Trajectory Segmented Consistency Distillation framework that pushes distilled diffusion to the extreme — generating usable images in just 1-2 steps. By combining consistency distillation with human feedback learning, it achieves the best quality-speed trade-off among distilled models.

Best For

1-step generationReal-time applicationsInteractive design toolsSpeed-critical workflowsPrompt experimentation

Prompting Tips

11-step mode produces usable results for rapid idea exploration
22-step and 4-step modes offer progressively higher quality
3Available in both SD 1.5 and SDXL variants
4Use the SDXL variant for higher resolution output
5Low CFG scale (1-1.5) works best with distilled steps
6Excellent for real-time prompt preview in creative tools

Syntax & Constraints

Compatible with standard SD/SDXL prompt syntax. Supports weighted prompts (element:weight) and negative prompts. Use very low CFG scale (1-1.5) and 1-4 steps.

Example Prompts

👤 Portrait Photography

Natural light portrait

(best quality:1.2), portrait photograph, woman with braided hair, soft natural window light, linen shirt, warm earthy tones, gentle expression, shallow depth of field, medium format film quality Negative: (low quality:1.4), distorted, bad anatomy

natural-lightwarmfilm

🏔 Landscape & Scenery

Tropical waterfall

(masterpiece:1.2), lush tropical waterfall in dense jungle, sunbeams filtering through canopy, emerald pool at base, mist and spray, vibrant greens, nature photography Negative: (low quality:1.4), dull, flat

tropicalwaterfalllush