Wonderful Launcher 文档ComfyUI Stable Diffusion 3.5 Guide: Setup and Workflows
How to run Stable Diffusion 3.5 in ComfyUI — model versions, download links, FP16 and FP8 workflows, and tips for best results.
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:
| Model | Parameters | Best For |
|---|---|---|
| SD 3.5 Large | 8B | Highest quality, professional use at 1MP resolution |
| SD 3.5 Large Turbo | 8B (distilled) | Fast generation in just 4 steps |
| SD 3.5 Medium | 2.5B | Consumer hardware, supports 0.25–2MP, easy to fine-tune |
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
| File | Location | Download |
|---|---|---|
| sd3.5_large_fp8_scaled.safetensors | models/checkpoints/ | HuggingFace |
| clip_g.safetensors | models/clip/ | HuggingFace |
| clip_l.safetensors | models/clip/ | HuggingFace |
| t5xxl_fp8_e4m3fn.safetensors | models/clip/ | HuggingFace |
If you have more than 32 GB of system RAM, use t5xxl_fp16.safetensors instead for better quality. The FP8 text encoder is for memory-constrained systems.
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:
- Load Checkpoint → sd3.5_large_fp8_scaled
- Triple CLIP Loader → clip_g, clip_l, t5xxl
- CLIP Text Encode (positive) → your prompt
- CLIP Text Encode (negative) → elements to avoid
- Empty Latent Image → 1024x1024 (recommended)
- KSampler → steps: 28, cfg: 4.5, sampler: euler, scheduler: sgm_uniform
- VAE Decode → Save Image
Setup: FP16 Original Version
For maximum quality with 16+ GB VRAM:
| File | Location | Download |
|---|---|---|
| SD 3.5 Large or Large Turbo checkpoint | models/checkpoints/ | Large / Large Turbo |
| clip_g.safetensors | models/clip/ | Same as above |
| clip_l.safetensors | models/clip/ | Same as above |
| t5xxl_fp16.safetensors | models/clip/ | HuggingFace |
If you previously used SD 3.0, you may already have the CLIP model files — they are shared across SD3 versions.
Key Differences from SD 1.5 / SDXL
| Aspect | SD 1.5 / SDXL | SD 3.5 |
|---|---|---|
| CLIP loader | Standard (1 or 2 CLIPs) | Triple CLIP Loader (3 CLIPs) |
| cfg value | 7–12 | 3.5–5.0 (lower is better) |
| Native resolution | 512 / 1024 | 1024x1024 (supports up to 2MP) |
| Scheduler | karras | sgm_uniform |
| Steps (Turbo) | N/A | 4 steps only |
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 chaircfg 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:
Related Guides
- Flux Guide — Alternative state-of-the-art image model
- Text to Image — Basic generation concepts
- LoRA Guide — Fine-tuning with LoRA adapters
ComfyUI SDXL Guide: Setup, Refiner Workflow & Best Practices
How to use Stable Diffusion XL (SDXL) in ComfyUI — model setup, base + refiner workflow, resolution tips, and ReVision for image-guided generation.
ComfyUI ControlNet Guide: Precise Control Over AI Image Generation
Learn what ControlNet is, how it works in ComfyUI, which control types exist, and how to set up your first ControlNet workflow step by step.