ComfyUI Stable Diffusion 3.5 Guide: Setup and Workflows

What is Stable Diffusion 3.5?

Stable Diffusion 3.5 is Stability AI's latest image generation model family, delivering significant improvements in image quality and prompt adherence over previous SD versions. It comes in three variants:

License

Free for research, non-commercial use, and organizations with annual revenue under $1 million. See the Community License for details.

Setup: FP8 Version (Recommended)

The FP8 version requires less VRAM and is simpler to set up.

Models

Workflow

The SD 3.5 workflow uses a Triple CLIP Loader (loads clip_g, clip_l, and t5xxl) instead of the standard single CLIP loader used by SD1.5:

1. Load Checkpoint → sd3.5_large_fp8_scaled
2. Triple CLIP Loader → clip_g, clip_l, t5xxl
3. CLIP Text Encode (positive) → your prompt
4. CLIP Text Encode (negative) → elements to avoid
5. Empty Latent Image → 1024x1024 (recommended)
6. KSampler → steps: 28, cfg: 4.5, sampler: euler, scheduler: sgm_uniform
7. VAE Decode → Save Image

Setup: FP16 Original Version

For maximum quality with 16+ GB VRAM:

Key Differences from SD 1.5 / SDXL

Tips for Best Results

Prompt style — SD 3.5 has excellent natural language understanding. Write descriptive sentences instead of keyword lists:

A cozy bookshop interior with warm lighting, wooden shelves filled
with colorful books, a tabby cat sleeping on a reading chair

cfg value — Keep it between 3.5 and 5.0. Higher values cause over-saturation and artifacts.

Large Turbo — When using the Turbo variant, set steps to exactly 4. More steps don't improve quality and waste time.

Common Issues and Fixes

Triple CLIP Loader node not found

• Update ComfyUI to the latest version — this node was added alongside SD 3.5 support

Out of memory

• Use the FP8 checkpoint and FP8 text encoder
• Reduce resolution to 768x768 for testing
• Close other GPU-intensive applications

Output is over-saturated or has color artifacts

• Lower cfg to 3.5–4.5
• This is the most common SD 3.5 mistake — it needs much lower cfg than SD 1.5

Image quality worse than expected

• Make sure you're using all three CLIP models (clip_g, clip_l, t5xxl)
• Try longer, more descriptive prompts — SD 3.5 benefits from natural language

Online Alternatives

If your hardware can't run SD 3.5 locally:

• SD 3.5 Large on HuggingFace Spaces
• SD 3.5 Large Turbo on HuggingFace Spaces
• Stability AI API

Related Guides

• Flux Guide — Alternative state-of-the-art image model
• Text to Image — Basic generation concepts
• LoRA Guide — Fine-tuning with LoRA adapters