Flux 2 Turbo accepts prompts up to standard token limits with configurable guidance_scale (default 2.5), six image_size presets or custom dimensions (512-2048px), and returns images with the seed used for generation. Structure prompts by front-loading subjects and use guidance values between 2.5-4.0 for most production scenarios.
Structuring Prompts for Flux 2 Turbo
Text-to-image diffusion models interpret natural language through learned associations between textual descriptions and visual features. Research demonstrates that optimized prompts significantly outperform unstructured user input in both automatic evaluation metrics and human preference ratings.1 Understanding how Flux 2 Turbo processes prompts directly influences the quality and consistency of generated outputs.
The model's architecture tokenizes and encodes prompts before guiding the diffusion process, meaning word order, specificity, and technical vocabulary all affect the final image. Flux 2 Turbo's rapid generation on fal makes iterative experimentation practical, allowing prompt refinement through multiple generations.
API Parameter Reference
The following table documents all available parameters for the Flux 2 Turbo text-to-image endpoint:
| Parameter | Type | Default | Description |
|---|---|---|---|
| prompt | string | required | Text description for image generation |
| image_size | enum/object | landscape_4_3 | Preset or custom dimensions (see below) |
| guidance_scale | float | 2.5 | Prompt adherence strength (0-20) |
| num_images | integer | 1 | Images per request (1-4) |
| seed | integer | random | Reproducibility control |
| enable_safety_checker | boolean | true | Content filtering |
| output_format | string | png | Output format: jpeg, png, webp |
Image Size Options
The image_size parameter accepts these presets:
- square_hd, square: Square aspect ratios
- portrait_4_3, portrait_16_9: Vertical orientations
- landscape_4_3, landscape_16_9: Horizontal orientations
For custom dimensions, pass an object with width and height (512-2048 pixels):
"image_size": {"width": 1024, "height": 768}
Response Schema
The API returns:
{
"images": [{ "url": "https://...", "content_type": "image/png" }],
"seed": 1234567890,
"prompt": "the prompt used"
}
The response includes the seed value used for generation. Store this seed to reproduce results or create variations with modified prompts.
falMODEL APIs
The fastest, cheapest and most reliable way to run genAI models. 1 API, 100s of models
Prompt Structure
Effective prompts share common structural elements: subject definition establishing what the image depicts, environmental context describing where the subject exists, style specifications indicating visual treatment, and technical attributes covering lighting and composition.
Consider the difference between "a vintage typewriter" and "A realistic photograph of a vintage typewriter with a sheet of paper inserted, sunlight falling across the desk." The second prompt specifies medium, includes contextual detail, and establishes atmosphere.
Front-Load Critical Details
Diffusion models assign greater weight to earlier tokens. Position your primary subject and essential attributes at the beginning.
Weak: "In a forest setting with morning light, a red fox standing on a moss-covered log"
Strong: "A red fox standing on a moss-covered log, forest setting, morning light filtering through trees"
Use Specific Descriptors
Abstract adjectives produce ambiguous results. Substitute vague terms with concrete visual descriptions.
Instead of "beautiful landscape": "Alpine valley at golden hour, snow-capped peaks reflecting warm sunset light, wildflower meadow in foreground"
Technical Terminology
Photography and art terminology provides established visual conventions:
- Lens: "50mm lens," "wide-angle shot," "macro photography"
- Lighting: "rim lighting," "soft diffused light," "dramatic chiaroscuro"
- Composition: "rule of thirds," "leading lines," "symmetrical composition"
- Medium: "oil painting," "watercolor," "digital illustration"
- Style: "impressionist style," "art deco," "Studio Ghibli aesthetic"
Guidance Scale Configuration
The guidance_scale parameter implements classifier-free guidance, which interpolates between conditional and unconditional generation to trade sample diversity against prompt fidelity.2
| Range | Behavior | Use Case |
|---|---|---|
| 1.0-2.0 | Creative interpretation | Artistic exploration, abstract concepts |
| 2.5-4.0 | Balanced adherence | Most production applications |
| 5.0-8.0 | Strict adherence | Technical illustrations, product renders |
| 9.0+ | Maximum adherence | Specific requirements only |
The default value of 2.5 provides reasonable results for general use. Values above 5.0 increasingly constrain interpretive flexibility, producing sharper details at the cost of natural variation.
Implementation Example
import fal_client
result = fal_client.subscribe(
"fal-ai/flux-2/turbo",
arguments={
"prompt": "Professional product photography of a luxury watch on black marble, studio lighting, macro lens, sharp focus",
"guidance_scale": 3.5,
"image_size": "square_hd",
"num_images": 2,
"seed": 42
}
)
# Access results
for image in result["images"]:
print(image["url"])
# Store seed for reproducibility
used_seed = result["seed"]
Seed-Based Iteration
To create variations while maintaining compositional consistency:
- Generate an initial image and note the returned seed
- Reuse that seed with modified prompts
- Compare outputs to isolate the effect of prompt changes
This technique supports A/B testing prompt modifications and developing consistent image series.
Prompt Examples
Product Photography
"Professional product photography of a luxury watch on black marble surface, studio lighting with soft shadows, macro lens, sharp focus on watch face, subtle reflections, commercial photography style"
Parameters: guidance_scale: 3.5, image_size: square_hd
Character Illustration
"Full body character design of a steampunk inventor, brass goggles on forehead, leather apron with tool pouches, mechanical arm prosthetic, confident pose, white background, concept art style"
Parameters: guidance_scale: 4.0, image_size: portrait_16_9
Architectural Visualization
"Modern minimalist house exterior, floor-to-ceiling windows, natural wood accents, native landscaping, late afternoon light, architectural photography, wide-angle perspective"
Parameters: guidance_scale: 3.0, image_size: landscape_16_9
Common Failure Modes
Contradictory Instructions
The model performs best with internally consistent descriptions. Avoid combining incompatible styles.
Problematic: "Photorealistic watercolor painting of a cartoon character"
Choose one direction and build supporting details around that aesthetic.
Missing Compositional Guidance
Without explicit framing, the model defaults to centered compositions. Include elements like "viewed from above," "low angle shot," or "off-center composition."
Ambiguous Color Descriptions
Terms like "colorful" lack specificity. Define palettes explicitly.
Vague: "colorful sunset scene"
Specific: "sunset scene with deep orange, magenta, and purple gradient sky, warm golden light on foreground"
Production Considerations
Batch Generation
Use num_images (1-4) to generate multiple variations per request. This reduces API calls when exploring prompt variations or selecting optimal compositions.
Error Handling
The API uses standard HTTP status codes. Implement exponential backoff for rate limits (429) and validate parameters client-side before submission to avoid 400 errors.
Cost
Flux 2 Turbo pricing is based on megapixels processed. A 1024x1024 image (1 megapixel) costs $0.008. Higher resolutions scale proportionally.
Further Resources
Access Flux 2 Turbo through fal's API. For complete endpoint documentation, see the API reference. The Queue API documentation covers asynchronous request handling for production integrations.
Recently Added
References
-
Hao, Y., Chi, Z., Dong, L., & Wei, F. (2023). Optimizing Prompts for Text-to-Image Generation. Advances in Neural Information Processing Systems (NeurIPS 2023). https://proceedings.neurips.cc/paper_files/paper/2023/file/d346d91999074dd8d6073d4c3b13733b-Paper-Conference.pdf ↩
-
Ho, J., & Salimans, T. (2022). Classifier-Free Diffusion Guidance. arXiv preprint arXiv:2207.12598. https://arxiv.org/abs/2207.12598 ↩
