Stable Diffusion 3.5 Local Deployment Complete Guide: Generate Unlimited Images for Free

Stable Diffusion 3.5 open-source makes "local unlimited image generation" a reality—text rendering, hands, and multi-subject scenes are significantly improved over SDXL. This article builds a complete local AI image workstation from scratch: hardware → installation → models → first image → advanced (LoRA/ControlNet/batch), and covers the Mac route and common errors.

1. Hardware Requirements

ConfigurationMinimumRecommendedFlagship

GPU VRAM8GB (for Medium/quantized)12-16GB24GB+ RAM16GB32GB64GB Storage50GB SSD100GB200GB+ (models take space) GPU ExamplesRTX 3060 12GRTX 4070/4080RTX 4090

SD 3.5 Large (8B): 16GB+ VRAM comfortable; quantized versions (GGUF/NF4) can reduce to ~10GB

SD 3.5 Medium (2.5B): 8-10GB VRAM works, excellent quality/speed ratio, recommended for beginners

Mac (Apple Silicon): Works (Metal/MPS), 32GB+ unified memory gives decent experience, speed ~1/3 to 1/5 of comparable NVIDIA; smoother Mac route is native apps like Draw Things

2. Installation Options Comparison

OptionDifficultyFeaturesBest For

ComfyUIMediumNode-based workflow, most powerful, fastest support for new modelsAdvanced / long-term users Stability MatrixEasyOne-click launcher managing multiple UIs and modelsThose who avoid environment setup SD.Next / ForgeEasy-MediumModern fork of A1111, WebUI styleThose used to traditional WebUI FooocusVery EasyOut-of-the-box but limited featuresQuick image generation only

Recommended: ComfyUI – Gets immediate support for new architectures like SD3.5/FLUX, workflows can be saved as JSON for reuse/sharing—this is the foundation for batch and custom services later.

3. ComfyUI Installation

Windows (Portable Package, Easiest)

Download the Portable package from ComfyUI's GitHub Releases (~1.5GB), extract to a path with only English characters

NVIDIA users run run_nvidia_gpu.bat, browser automatically opens 127.0.0.1:8188

Recommended: Install ComfyUI-Manager plugin (git clone into custom_nodes/) – it can one-click install any missing nodes/models later

Mac / Linux

bash
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI
pip install -r requirements.txt
python main.py            # Mac uses MPS automatically; add --lowvram if VRAM is tight

4. Download Models (Critical Path)

SD 3.5 weights are on Hugging Face in the stabilityai repository (requires accepting the community license; free for personal and small/medium commercial use, large enterprises need commercial license – refer to Stability's official terms):

sd3.5_medium.safetensors → place in ComfyUI/models/checkpoints/

Three text encoders clip_l, clip_g, t5xxl_fp8 → place in ComfyUI/models/clip/ (SD3 architecture has separate text encoders; missing t5 is the #1 beginner error)

ComfyUI's official example workflow (drag an official example image into the window to auto-load the node graph) is the fastest way to start.

5. First Image: Parameter Tuning

Steps: 25-30 steps are sufficient for SD3.5

CFG: 4-6 (lower than SDXL's 7-8; too high causes "burning" – see the classifier-free guidance section in Diffusion Models Technical Deep Dive)

Prompt: SD3.5's T5 encoder handles natural language long sentences well; write complete English descriptive sentences for better results than stacking tags; negative prompts have less effect than in SD1.5, simply use blurry, lowres

Resolution: Native 1MP range (1024×1024, 1152×896, etc.), don't directly go to 2K – first generate at 1MP then upscale

6. Advanced Roadmap

LoRA: Community style models (search SD3.5 tag on Civitai) place in models/loras/, add a LoRA Loader node in the workflow; train your own character/style LoRA with kohya_ss or OneTrainer, starting from 20-40 source images

ControlNet: Pose/lineart/depth control for composition – core capability for custom orders

Upscaling: SD upscale workflow (model upscale + resampling) is much sharper than simple interpolation

Batch Generation: ComfyUI has a built-in API (/prompt endpoint sends JSON workflow), combined with scripts for batch image generation – this powers the POD/stock image route in AI Illustration Monetization

FLUX vs SD3.5? FLUX dev has slightly higher image quality ceiling, ecosystem and commercial licensing differ – see SD vs FLUX Comparison

7. Common Error Quick Reference

SymptomCause/Solution

Output is all noise/blackText encoders not fully installed (add t5xxl); or precision issue, add --force-fp16 CUDA out of memorySwitch to Medium/quantized version; add --lowvram on startup; close other VRAM-heavy browser tabs Extremely slowConfirm running on GPU (check startup logs); Windows: check if PyTorch CPU version was installed Node highlighted red, missingComfyUI-Manager → Install Missing Custom Nodes

FAQ

Q: Local deployment vs Midjourney? Local = zero marginal cost, privacy, controllability (LoRA/ControlNet), batch capability; MJ = stunning out-of-the-box, zero maintenance. For client work, know both: MJ for concepts, local for controllable delivery.

Q: What if I don't have an NVIDIA GPU? Mac: use MPS or Draw Things; AMD cards on Linux with ROCm work but are fiddly; or rent cloud GPUs (AutoDL/RunPod, hourly billing) for batch tasks.

Q: Is this the same as local LLM deployment? Similar concept (open-source weights + local inference), different toolchain – for LLMs see Ollama vs LM Studio vs Jan.

---

*Last updated: June 2026. Model licenses and tool versions change rapidly; refer to official repositories.*