FLUX.2 and Stable Diffusion 3.5 — Open source
Explore the world of open-source image models: FLUX.2 from Black Forest Labs and Stable Diffusion 3.5. Learn to use them through Hugging Face, APIs, and locally with ComfyUI.
🌊 FLUX.2 — The Open-source Giant
FLUX.2, developed by Black Forest Labs (founded by former Stable Diffusion creators), is the most advanced open-source image model available in 2026.
FLUX.2 dev — The main model
32 billion parameters. Quality comparable to Midjourney V7 in many scenarios. Supports resolutions up to 4 megapixels. Ideal for production at scale via API or local use with a powerful GPU (16GB+ VRAM recommended for development).
FLUX.2 klein — Ultra-fast
An optimized version that generates images in under 1 second on GPUs with just 8GB VRAM. Perfect for rapid prototyping and local use on modest hardware. Slightly lower quality than dev, but unmatched speed.
💡 Why FLUX.2 Matters
With FLUX.2, you have full control: no monthly fees, no generation limits, no restrictive terms of use, and you can train the model on your own data. For high-volume commercial production, the cost per image is practically zero (just electricity/GPU).
🎯 Stable Diffusion 3.5
Stability AI remains relevant with SD 3.5, which introduced the new MMDiT (Multimodal Diffusion Transformer) architecture:
| Variant | Parameters | Minimum VRAM | Best use |
|---|---|---|---|
| SD 3.5 Large | 8B | 12GB | Maximum quality, professional use |
| SD 3.5 Medium | 2.6B | 8GB | Quality/speed balance |
FLUX.2 vs. SD 3.5: which should you choose?
FLUX.2 generally produces higher-quality images and is faster with FLUX klein. SD 3.5 has the advantage of a more mature ecosystem (more extensions, LoRAs, and community support). For beginners, we recommend starting with FLUX.2 via Hugging Face and exploring SD 3.5 when you want more control with ComfyUI.
🚀 How to Use: Local vs API vs Hugging Face
1. Hugging Face Spaces (easiest)
Access it directly in your browser without installing anything. Go to huggingface.co/spaces and search for "FLUX" or "Stable Diffusion". Enter your prompt and generate. There may be a queue during peak hours, but it’s 100% free.
2. API (for integration)
Use the Hugging Face, Replicate, or Black Forest Labs API directly. Ideal for automations with n8n or scripts. Pay per use (a few cents per image), with no queues and fast responses.
3. Local with ComfyUI (maximum control)
Install ComfyUI on your computer, download the model, and have full control. Requires an NVIDIA GPU with 8GB+ VRAM. No recurring costs, no limits, no queues. A steeper learning curve, but rewarding.
🔧 ComfyUI for Beginners
ComfyUI is a node-based visual interface for running diffusion models locally. It may seem complex at first, but the concept is simple:
Basic ComfyUI workflow
- 1. Loader — Loads the model (FLUX.2 or SD 3.5)
- 2. CLIP Text Encode — Converts your prompt into embeddings
- 3. KSampler — Generates the image (diffusion process)
- 4. VAE Decode — Converts the result into a visible image
- 5. Save Image — Saves the result
⚠️ For this course
ComfyUI will be explored in greater depth in Levels 2 and 3. At this level, the focus is on using Hugging Face Spaces and APIs to generate images with FLUX.2. If you have a powerful GPU and want to get ahead, consult the official ComfyUI documentation on GitHub.
✅ Lesson Checklist
- I understand the differences between FLUX.2 dev and klein
- I know the Stable Diffusion 3.5 variants
- I know the 3 ways to use open-source models (HF, API, local)
- I generated at least 3 images via Hugging Face Spaces
- I understand the basic concept of ComfyUI