If you are running image generation locally in 2026, you have two real choices: FLUX (from Black Forest Labs) or the Stable Diffusion family (from Stability AI and the open source community). Both run on consumer GPUs, both have active ecosystems, and both produce genuinely good images. But they excel at very different things.
I have been running both for six months across various hardware. Here is the honest comparison, including when each wins and the specific hardware you need.
The Contenders
FLUX family:
- FLUX.2 [dev]: 32B parameters, non-commercial license, highest quality
- FLUX.2 [schnell]: 32B parameters, Apache 2.0, optimized for speed (4 steps)
- FLUX.2 [klein] 4B: 4B parameters, Apache 2.0, fits on 8GB GPUs
Stable Diffusion family:
- Stable Diffusion 3.5 Large: 8B parameters, improved architecture
- Stable Diffusion 3.5 Medium: 2.5B parameters, fast and accessible
- SDXL 1.0: 6.6B parameters (base + refiner), massive community ecosystem
- SD 1.5: legacy but still relevant for specific LoRA workflows
For the API-based alternatives when local is not enough, check my image generation API pricing breakdown. And if you want to integrate image generation into an app via API, my Python image generator tutorial covers that workflow end to end.
Hardware Requirements
This is where the decision starts for most people. Your GPU dictates which models you can even run.
8GB VRAM (RTX 4060, RTX 3070):
- FLUX.2 [klein] 4B: runs well, 4 steps, good quality for the size
- SDXL: runs with aggressive optimization (fp16, tiled VAE)
- SD 3.5 Medium: comfortable fit, fast generation
- SD 1.5: runs easily with headroom for large batches
12GB VRAM (RTX 4070, RTX 3060 12GB):
- FLUX.2 [dev/schnell]: runs with fp8 quantization, slower but works
- SD 3.5 Large: comfortable fit at fp16
- SDXL: runs natively with room for LoRAs and ControlNet
16GB VRAM (RTX 4080, RTX 5060 Ti):
- FLUX.2 [dev/schnell]: runs well at fp16, good speed
- All SD models: run comfortably with multiple ControlNets loaded
24GB VRAM (RTX 4090, RTX 3090):
- FLUX.2 [dev]: full quality, fast generation, best results
- Everything else: runs without any compromises
For a deeper dive into VRAM and model compatibility, see my VRAM requirements guide. If you want the complete setup guide for running FLUX locally, I have a dedicated tutorial.
Quality Comparison
Let me be opinionated here: FLUX produces better images for most use cases. But “better” needs qualification.
FLUX wins at:
- Photorealism (skin texture, lighting, reflections)
- Text rendering (signs, labels, UI mockups with readable text)
- Anatomical accuracy (hands, fingers, complex poses)
- Following complex prompts with multiple subjects
- Consistent composition and framing
Stable Diffusion wins at:
- Artistic and stylized outputs (anime, illustration, concept art)
- LoRA ecosystem (thousands of fine-tuned styles available)
- Specific character consistency (with trained LoRAs)
- Speed per image on equivalent hardware
- Mature inpainting workflows
- ControlNet variety (dozens of preprocessors, well-tested)
For production-quality photorealistic images (product shots, marketing materials, blog hero images), FLUX is the clear winner. For creative and artistic work, the SD ecosystem offers more variety and control through LoRAs.
Speed Benchmarks
Tested on an RTX 4090 24GB with ComfyUI, generating 1024x1024 images:
| Model | Steps | Time per Image | Images per Minute |
|---|---|---|---|
| FLUX.2 [dev] fp16 | 20 | 8.2 sec | 7.3 |
| FLUX.2 [schnell] fp16 | 4 | 1.9 sec | 31.6 |
| FLUX.2 [klein] 4B | 4 | 0.9 sec | 66.7 |
| SD 3.5 Large fp16 | 28 | 4.1 sec | 14.6 |
| SDXL (base only) | 30 | 3.8 sec | 15.8 |
| SD 1.5 | 20 | 1.2 sec | 50.0 |
On a 12GB GPU (RTX 4070), FLUX.2 dev with fp8 quantization takes about 22 seconds per image. SDXL on the same card takes 5.5 seconds. That is a 4x speed difference that matters for iterative workflows where you are generating dozens of variations.
The LoRA Ecosystem
This is Stable Diffusion’s biggest advantage and it is not close.
SDXL LoRAs available on CivitAI: 50,000+ SD 1.5 LoRAs: 100,000+ FLUX LoRAs: ~3,000 (growing but still small)
LoRAs let you fine-tune style, character appearance, specific objects, and artistic techniques without training a full model. For anyone doing consistent character work, brand-specific styles, or niche artistic outputs, the SD LoRA library is irreplaceable.
FLUX LoRAs exist and work well, but the ecosystem is 12 to 18 months behind SD in variety. If you need a very specific style (like “1990s anime cel shading” or “watercolor botanical illustration”), SD has a LoRA for it. FLUX probably does not yet.
ComfyUI Workflows
Both models work excellently with ComfyUI, which has become the standard local generation interface.
FLUX in ComfyUI:
- Native support since mid-2024
- Excellent text-to-image workflows
- Image-to-image and inpainting work but fewer community workflows
- IP-Adapter support available (style transfer)
- Upscaling pipelines well-established
Stable Diffusion in ComfyUI:
- Extremely mature, thousands of shared workflows
- ControlNet (pose, depth, edge, normal map, etc.) fully supported
- AnimateDiff for video generation
- Regional prompting for multi-subject control
- Inpainting workflows are battle-tested
- InstantID and IP-Adapter for face consistency
If you are building complex multi-step pipelines (generate, inpaint background, add specific subject, upscale), SD in ComfyUI has more tested workflows and community support.
My Recommendation by Use Case
Choose FLUX if you need:
- Product photography for e-commerce
- Marketing materials and ad creatives
- Blog and social media images (photorealistic)
- UI/UX mockups (text rendering matters)
- Anything where photorealism is the goal
- Quick generation with [klein] for iterations
Choose Stable Diffusion if you need:
- Artistic and stylized content
- Consistent characters across many images
- Specific trained styles (via LoRAs)
- Video generation (AnimateDiff)
- Complex multi-step generation pipelines
- Maximum speed on limited hardware
Choose both if:
- You have 24GB VRAM and disk space (both fit)
- You do varied creative work
- You want the best tool for each specific task
I run both. FLUX handles my blog images and product shots. SDXL handles anything requiring specific artistic styles or character consistency. This is the setup I describe in my local-first AI developer stack. If you want to start with a zero-cost approach, ComfyUI with SDXL runs fine on modest hardware.
Practical Setup Guide
Installing both via ComfyUI:
- Install ComfyUI (one installation supports all models)
- Download FLUX.2 dev checkpoint (~24GB) to
models/unet/ - Download FLUX text encoders (T5-XXL + CLIP-L) to
models/clip/ - Download FLUX VAE to
models/vae/ - Download SDXL base checkpoint (~6.5GB) to
models/checkpoints/ - Download any SDXL LoRAs to
models/loras/
Total disk space: approximately 60GB for both full setups.
For model management and switching, Ollama handles LLMs but ComfyUI is the equivalent for image models. Keep both running and switch based on task.
Cost Analysis: Local vs API
Running locally means zero marginal cost per image after hardware. Let me quantify the savings:
Local (RTX 4090):
- Hardware amortization: $44/month
- Electricity for image generation (2 hours/day): $5/month
- Cost per image: effectively $0 at volume
- 1,000 images/month: $0.05/image (amortized)
API alternatives:
- fal.ai FLUX: $0.01 to $0.04 per image
- Replicate FLUX: $0.003 to $0.05 per image
- Stability API: $0.02 to $0.06 per image
At 100 images/month, APIs are cheaper (no hardware cost). At 500+ images/month, local wins decisively. At 1,000+ images/month, local is 10x cheaper.
If you are generating fewer than 200 images monthly and do not need real-time generation, APIs make more financial sense. My image generation API pricing comparison covers all the options. The same cost optimization principles I outlined in my guide to reducing AI API costs apply here: cache results, batch processing, and use the cheapest model tier that meets your quality bar.
Future Outlook
FLUX is evolving faster than Stable Diffusion in 2026. Black Forest Labs has stronger funding and a clear roadmap. Stability AI has had leadership changes and financial pressure, though the open source community around SD remains massive and productive.
My prediction: FLUX will dominate production and commercial use cases within 12 months. SD will remain the creative and community tool of choice because of its LoRA ecosystem and permissive licensing (SDXL is open). Both will coexist for years.
For developers building AI image generation into products, consider building with APIs first and moving to local inference only when volume justifies the hardware.
FAQ
Can I run FLUX on an 8GB GPU?
Yes, but only FLUX.2 [klein] 4B, which is a much smaller model. The full FLUX.2 dev 32B model needs 12GB minimum (with fp8 quantization and compromises) and runs best on 24GB. If you have 8GB, SDXL or SD 3.5 Medium give better quality-per-VRAM than FLUX klein, though klein is fast and decent for quick iterations.
Which model produces better hands and anatomy?
FLUX wins decisively on anatomical accuracy. The 32B parameter count and newer training data give it much better understanding of human anatomy compared to SDXL. If your use case involves people (portraits, lifestyle photography, product modeling), FLUX reduces the rate of anatomical errors from roughly 30% (SDXL) to under 5%.
Is Stable Diffusion dead in 2026?
Absolutely not. SDXL has the largest community of any open image model, with tens of thousands of LoRAs, ControlNet models, and workflows. SD 3.5 brought real improvements in quality. The ecosystem momentum means SD will remain relevant for years regardless of what Black Forest Labs ships. For artistic and stylized use cases, SD arguably still leads.
How much disk space do I need for both models?
Plan for 60 to 80GB total. FLUX.2 dev is about 24GB (checkpoint + text encoders + VAE). SDXL base is about 6.5GB. Add 10 to 20GB for LoRAs, ControlNet models, and upscalers. If disk space is tight, pick one and commit. I recommend FLUX for most developers since photorealistic content is more commonly needed for professional work.
Should I use FLUX or SD for generating AI video frames?
For video workflows (frame interpolation, AnimateDiff-style generation), Stable Diffusion is the only practical choice in 2026. AnimateDiff was built for SD architectures and has no FLUX equivalent. FLUX excels at single high-quality frames, but SD has the video pipeline maturity. For dedicated AI video, see my video API pricing guide for cloud options.