Stability AI's SD 3.5 series shipped in late 2024 and matured through 2026 into a genuinely production-capable open-weights image stack — especially when paired with ComfyUI's node-based workflow engine. For teams generating thousands of images per month and willing to invest in local infrastructure, SD 3.5 + ComfyUI is the cheapest and most flexible option. Here is the practical guide.
What changed in 2026
- SD 3.5 Large + Turbo + Medium cover the spectrum — 8B params for highest quality (Large), 2.5B for speed (Medium), 4-step Turbo for real-time UX.
- ComfyUI Manager matured into the most-used SD interface, with thousands of community nodes.
- Permissive Stability community license allows commercial use up to $1M annual revenue without paid license — covers most teams.
Hardware requirements
| Model |
VRAM |
Speed (RTX 4090) |
| SD 3.5 Medium |
8GB |
~2s/image |
| SD 3.5 Large |
16GB |
~5-8s/image |
| SD 3.5 Large + ControlNet |
20GB |
~10-15s/image |
For Mac users: M3 Pro / M4 Pro with 32GB+ unified memory runs SD 3.5 Large; M3 Max / M4 Max gives ~3x speed.
ComfyUI primer
ComfyUI is a node-graph interface for image generation pipelines. Each node does one thing (load model, encode prompt, sample, decode, save). Connect them; build complex pipelines.
A basic SD 3.5 workflow has ~5 nodes:
- Load Checkpoint — pick SD 3.5 Large
- CLIP Text Encode (positive + negative prompts)
- Empty Latent Image — set dimensions
- KSampler — model + prompts + latent + steps + cfg
- VAE Decode + Save Image
Save the workflow as JSON; reuse and modify.
Recommended ComfyUI setup
Install via Pinokio (one-click installer) or Comfy.org's official installer. Then install:
- ComfyUI-Manager (essential for installing more nodes)
- rgthree-comfy (better UX, multi-image preview)
- ComfyUI-AdvancedLivePortrait for character consistency
- ControlNet Aux for Canny, Depth, Pose preprocessing
SD 3.5 Large workflow patterns
Photorealistic: cfg 4.5, steps 28, sampler dpmpp_2m. Negative prompt: low quality, jpeg artifacts, watermark, low resolution.
Stylized: cfg 5.5, steps 35, sampler dpmpp_3m_sde. Add style LoRA at strength 0.8.
Fast iteration: SD 3.5 Turbo, 4 steps, cfg 1.5, sampler dpm_sde_gpu. Real-time on RTX 4090.
LoRA training for brand-specific styles
LoRAs (low-rank adaptations) are small (~50-200MB) trainable add-ons. Train one on 30-50 images of your brand style; produce on-brand output reliably. Tools: kohya-ss, OneTrainer, or ComfyUI's built-in trainer.
Training time: 1-2 hours on RTX 4090 for a quality LoRA.
SD 3.5 vs Flux vs Midjourney
| Aspect |
SD 3.5 Large |
Flux Dev (local) |
Midjourney v7 |
| Quality (1-10) |
7.5 |
8.5 |
9.0 |
| Prompt adherence |
7.0 |
9.0 |
7.0 |
| Text rendering |
6.0 |
8.5 |
5.0 |
| Cost per 1000 images (local) |
$0 |
$0 |
n/a |
| Style flexibility |
High (many LoRAs) |
Medium |
High |
For sheer image quality, Flux Dev usually wins. SD 3.5 wins on flexibility, ecosystem (10x more LoRAs and plugins), and zero per-image cost once you've set up local hardware.
When local makes sense
>500 images/mo: local hardware pays back in 6-12 months vs API costs.
Sensitive content: legal, medical, brand-internal imagery you don't want flowing through third-party APIs.
Heavy customization: trained LoRAs, custom ControlNet variants, complex workflows. Cloud APIs typically don't expose these.
Skip local if: you generate <200 images/mo. API services like Replicate, fal, or Together cost $5-25/mo for that volume.
FAQ
Can I use SD 3.5 commercially?
Yes if your annual revenue is under $1M. Above that, Stability requires a license (~$2K/mo for $1M-$10M revenue).
What about ControlNet for SD 3.5?
ControlNets for SD 3.5 are available for Canny, Depth, Pose, and Tile via Stability and community releases.
Is SDXL still relevant?
Less so in 2026. SD 3.5 Large beats SDXL on most metrics. SDXL still has a richer LoRA ecosystem (more legacy options).
Where to go next
For related guides see Midjourney v7 prompting guide, Flux Pro prompting guide, and Local LLM setup guide for 2026.