Using Flux.1 [dev] Locally for Commercial Use! How to Use Fal.ai's API Node

Introduction

Hello, I'm Easygoing.

This time, we'll explore how to create illustrations for commercial use with the Flux.1[dev] model.

Flux.1[dev]: A High-Quality Photorealistic Model

Flux.1[dev] is an AI model that emerged in August 2024, capable of generating high-quality photorealistic illustrations.

Download numbers for major photorealistic models on Hugging Face over the past month

Although about a year has passed since its debut, Flux.1[dev] continues to enjoy broad support as a high-quality photorealistic model.

Two Types of Licenses for Flux.1[dev]

Flux.1[dev] comes with two types of licenses:

Flux.1[dev] Non-Commercial License

The Flux.1[dev] model distributed for free on Hugging Face is under a non-commercial research license.

The Flux.1[dev] Non-Commercial License has had ambiguous provisions regarding the commercial use of outputs. Initially, it included wording that could be interpreted as allowing commercial use of outputs, but on June 28, 2025, this section was removed in a license revision, only to be reinstated the next day following user inquiries.

This section remains (likely intentionally) ambiguous, and Black Forest Labs, the company behind Flux.1[dev], may not explicitly permit commercial use of outputs, so caution is advised.

For delivering illustrations for commercial purposes, it’s safer to avoid using the Flux.1[dev] Non-Commercial License.

Commercial License for Service Providers

In contrast, the license starting at $999/month is designed for AI service providers. Platforms like Fal.ai and Replicate offering Flux.1[dev] image generation services explicitly allow commercial use of outputs.

When delivering illustrations for commercial use, this commercial license ensures you can use Flux.1[dev] outputs with confidence.

Using Both Licenses!

When creating AI illustrations, you often need to generate multiple images and select the best ones. Generating everything with the paid commercial license can be inefficient.

So, I devised a method to use both licenses.

Specifically, I use the non-commercial license for the “gacha” phase—adjusting prompts or seeds—and once a satisfactory illustration is achieved, I use the same prompt and seed with the commercial license to generate the final illustration.

Workflow!

Here’s the workflow I used:

• Fal_API_20250726.json

The upper-right part of the workflow represents standard local Flux.1[dev] generation.

The lower-right part uses the Fal API Flux Dev custom node to run Flux.1[dev] online via Fal.ai’s API.

Standardizing on the Original Flux.1[dev] Model

To ensure consistent results between local and online generation, I standardized on the original Flux.1[dev] model, including the text encoder.

Models Used

• Text Encoder
  • t5xxl_v1_1
  • clip-l
• Flux.1[dev] Model
  • flux1-dev
• VAE
  • ae.safetensors

Since online services likely use FP8 format to save VRAM and computational resources, I also configured the local setup to use FP8 format in ComfyUI.

ComfyUI Launch Settings

• --fp8e4m3fn-unet
• --fp8e4m3fn-text-enc

Comparing Actual Illustrations!

Let’s compare illustrations generated locally and online.

City Nightscape

Woman in a White Shirt

Office with Greenery

Online Flux.1[dev] Has Lower Image Quality!

Online Flux.1[dev] illustrations are significantly lower in quality compared to local ones.

Online images tend to have simplified, near-symmetrical compositions, fewer details, and a hazy, whitish appearance.

This is a phenomenon observed in previous tests when computational precision is reduced.

Both local and online setups used the same model and FP8 format, so why the difference?

Online Generation Is Extremely Fast!

Comparing generation times, local generation takes 1–2 minutes per image, while online generation completes in just a few seconds.

Fal.ai’s GPU service, primarily for business use, provides NVIDIA H100 GPUs, which are about three times more powerful than the consumer-grade RTX 4090. Even so, the generation speed is remarkably fast.

Is Online Using Full Optimization Techniques?

Image generation AI employs various optimization techniques, such as caching, to speed up processing.

Caching (WaveSpeed)

Caching (TeaCache) and Sage Attention

These techniques reduce computational load and speed up generation but often come at the cost of lower image quality, and controlling them is challenging.

I suspect online generation uses these optimization techniques extensively to reduce computational load.

Even with a 20-step setting, online generation sometimes produces clearly incomplete illustrations, likely due to caching, making it difficult to achieve local-level quality.

Comparing with the Highest-Precision Local Model!

Now, let’s compare online results with illustrations generated locally using the highest-precision model, with the same prompt and seed.

• Text Encoder
  • flan_t5_xxl_TE-only-FP32
  • CLIP-SAE-ViT-L-14-FP32
• Flux.1[dev] Model
  • Flux.1[dev]

Anime-Style Woman in a Kimono

Woman in a Black Shirt

Local illustrations not only have higher quality but also depict characters with more convincing presence.

High-precision models and text encoders seem to capture not just the words in the prompt but also the contextual meaning, translating it into the illustration.

How to Use Fal.ai’s API

Here’s how I used Fal.ai’s API.

Signing Up for Fal.ai

To use Fal.ai, you need to sign up.

You can sign up using GitHub, Google, or an enterprise domain.

Upon registration, you receive $1.00 in free credit.

Issuing an API Token

After signing up, go to the API Keys tab from the homepage and click “Add key.”

Enter a name for the key, click “Done” to create it, and copy the API key.

Installing the ComfyUI-Fal-API-Flux Custom Node

Next, launch ComfyUI locally and install the ComfyUI-Fal-API-Flux custom node.

After installation, a config.ini file is created in the custom_nodes\comfyui-fal-api-flux folder. Paste the API key into this file:

[falai]
api_key = 🔵🔵🔵🔵🔵

Then, run the workflow to generate Flux.1[dev] illustrations online.

API usage fees vary by computational load but are generally around $0.025 per run, which is very reasonable.

Be sure not to share the API key or config.ini file with others.

Conclusion: Online Quality Is a Challenge

• Fal.ai’s API enables commercial use of Flux.1[dev].
• Online generation is fast.
• Quality remains a challenge.

The Flux.1[dev] model excels in photorealistic illustrations, but its licensing is often debated.

This time, I aimed to create high-quality, commercially usable illustrations using Fal.ai’s Flux.1[dev], but unfortunately, I couldn’t replicate local quality.

It would be ideal if online Flux.1[dev] allowed choosing compression formats or optimization settings, but such options are currently unavailable.

I’ll continue exploring ways to generate high-quality illustrations for commercial use.

Thank you for reading!