Flux AI Image Generation Guide 2025: FLUX.1 Models & Setup
FLUX.1 is the leading open-weight image generation family in 2025 — created by Black Forest Labs (ex-Stability AI team). Choose schnell (free, MIT), dev (quality, non-commercial), or pro (best quality, API-only). Run via Replicate API, locally with ComfyUI, or free on HuggingFace Spaces.
1. What is Flux AI?
Black Forest Labs launched FLUX.1 in August 2024. The founding team includes Robin Rombach and other key researchers behind Stable Diffusion. Unlike SD XL, Flux uses a hybrid transformer-diffusion architecture (DiT) that processes text and image tokens together, leading to better text rendering and more literal prompt following.
Key features
2. FLUX.1 model variants
There are three FLUX.1 models. The right one depends on your speed, quality, and licensing needs.
| Model | Speed | Quality | API Price | License |
|---|---|---|---|---|
| FLUX.1 schnell | Fastest (4 steps) | Good | ~$0.003/img (fal.ai) | MIT — free commercial |
| FLUX.1 dev | Medium (28 steps) | High | $0.025/img (Replicate) | Non-commercial only |
| FLUX.1 pro | Slower (API only) | Best | $0.055/img (Replicate) | API-only, no self-host |
3. Using Flux via Replicate API
Replicate hosts all three FLUX.1 variants behind a simple REST API. Get a token at replicate.com, add $5 credit, and you can generate images programmatically without any local GPU setup.
Quick test with curl (schnell)
curl -s -X POST \
-H "Authorization: Bearer $REPLICATE_API_TOKEN" \
-H "Content-Type: application/json" \
-d '{"input": {"prompt": "A photorealistic mountain lake at golden hour, 8k, sharp focus"}}' \
https://api.replicate.com/v1/models/black-forest-labs/flux-schnell/predictions The response includes a prediction ID. Poll GET /v1/predictions/{id} until status is succeeded, then retrieve the output URL.
Python example (dev model)
Install: pip install replicate
import replicate
output = replicate.run(
"black-forest-labs/flux-dev",
input={
"prompt": "A photorealistic mountain lake at golden hour, 8k, sharp focus",
"width": 1024,
"height": 1024,
"num_outputs": 1,
"guidance_scale": 3.5,
"num_inference_steps": 28,
}
)
print(output[0]) # URL to generated image 4. Using Flux locally with ComfyUI
ComfyUI is the best tool for running Flux locally — it provides a node-based workflow editor that gives you full control over the generation pipeline, LoRA loading, and batching.
Hardware requirements
| Model | Min VRAM | Notes |
|---|---|---|
| FLUX.1 schnell (fp8) | 8GB VRAM | Fast, MIT license, great for iteration |
| FLUX.1 schnell (fp16) | 16GB VRAM | Full precision, best schnell quality |
| FLUX.1 dev (fp8) | 12GB VRAM | Recommended for local high-quality |
| FLUX.1 dev (fp16) | 24GB VRAM | Full precision, RTX 3090/4090, A100 |
Install ComfyUI
Clone the ComfyUI repo and install requirements: pip install -r requirements.txt. On Windows, use the standalone installer package from github.com/comfyanonymous/ComfyUI.
Download FLUX.1 weights from HuggingFace
For schnell: black-forest-labs/FLUX.1-schnell (no auth required). For dev: black-forest-labs/FLUX.1-dev (requires HF account + accepting license). Place the .safetensors file in ComfyUI/models/unet/.
Download the T5 and CLIP text encoders
Flux uses two text encoders: CLIP-L and T5-XXL. Download from comfyanonymous/flux_text_encoders on HuggingFace. Place in ComfyUI/models/clip/.
Load the Flux workflow in ComfyUI
ComfyUI includes example Flux workflows in the examples/ folder. Open flux_dev_example.json or flux_schnell_example.json. Enter your prompt in the CLIPTextEncode node and click Queue Prompt.
5. Free Flux on HuggingFace Spaces
The fastest way to try Flux without any setup is through HuggingFace Spaces. No GPU required on your end — HuggingFace runs the model in the cloud.
Available Spaces
No login required. Fast generation (seconds). MIT license — images you generate can be used commercially.
Requires HF account + license acceptance. Higher quality than schnell. Non-commercial use only.
Many community Spaces offer Flux with additional controls (aspect ratio, seed, steps). Search "FLUX" on HuggingFace Spaces to find them.
6. Prompt tips for Flux
Flux understands natural language much better than older Stable Diffusion models. You don't need to write comma-separated tags — write sentences instead.
Natural language over tags
Old Stable Diffusion style (less effective with Flux)
Flux-native natural language style (works better)
Effective Flux prompt patterns
Copy-paste starter prompts
7. Flux vs Midjourney vs DALL-E 3 vs SD XL
How FLUX.1 compares to the major AI image generation tools in 2025.
| Tool | Quality | Speed | Cost | Control | Self-host |
|---|---|---|---|---|---|
| FLUX.1 pro | ★★★★★ | Medium | $0.055/img | High | No |
| Midjourney V6 | ★★★★★ | Medium | $10/mo (200 imgs) | Low | No |
| DALL-E 3 | ★★★★ | Medium | $0.04–$0.12/img | Medium | No |
| FLUX.1 dev | ★★★★ | Medium | $0.025/img or free (local) | Very high | Yes (non-commercial) |
| FLUX.1 schnell | ★★★★ | Very fast | ~$0.003/img or free | Very high | Yes (MIT license) |
| SD XL | ★★★ | Fast | Free (local) | Very high | Yes (Apache 2.0) |
8. Commercial licensing
Understanding what you can do commercially with each Flux variant is critical before using generated images in products or client work.
FLUX.1 schnell — MIT License (Free Commercial)
Schnell is MIT-licensed: you can use generated images in commercial products, sell them, use them in client work, incorporate into apps, and redistribute without restriction. You must include attribution in derivative works of the model weights themselves (not the images).
FLUX.1 dev — Non-Commercial License
Dev weights have a non-commercial license: you can use them for personal projects, research, and experimentation, but not in commercial products, client deliverables, or any revenue-generating context. Using dev via Replicate/fal.ai API still falls under this restriction.
FLUX.1 pro — API Commercial Use (check terms)
Pro is available only via Replicate and fal.ai APIs. Images generated through these APIs are covered by each platform's terms — generally allowing commercial use of outputs. Check Replicate's and fal.ai's terms of service for the specific usage rights granted for generated images.
Monitor Replicate, fal.ai and HuggingFace status at Prismix
Flux generation failures are often platform outages, not model errors. Check Replicate and HuggingFace status at Prismix before debugging your prompts.
FAQ
What is FLUX AI image generation?
FLUX is a family of open-weight image generation models from Black Forest Labs (ex-Stability AI researchers). The FLUX.1 lineup includes schnell (fastest, MIT-licensed), dev (higher quality, non-commercial), and pro (best quality, API-only). Flux uses a hybrid transformer-diffusion architecture (DiT) that understands natural language better than tag-based SD models.
Is FLUX free to use?
FLUX.1 schnell is MIT-licensed — completely free for personal and commercial use locally. FLUX.1 dev is free to download but non-commercial. FLUX.1 pro requires a paid API (Replicate $0.055/img). HuggingFace Spaces offers free schnell and dev access (with queue waits).
Can I run Flux locally?
Yes — schnell and dev run locally with ComfyUI. You need at least 8GB VRAM for schnell (fp8) or 12GB+ for dev. Apple Silicon (M1-M4) is supported via MPS backend. CPU-only is very slow (30+ min/image).
Flux vs Midjourney: which is better?
FLUX.1 pro matches Midjourney V6 in photorealism and beats it in text rendering. Midjourney wins on artistic style and its community. Flux wins on API access, local self-hosting, MIT license (schnell), and natural language instruction-following.