FLUX-2

flux-2-turbo-edit is an advanced image-to-image editing model developed by Black Forest Labs as part of the FLUX.2 family. It specializes in ultra-fast, precise modifications to input images using natural language instructions and hex color control, enabling targeted edits like changing weather, backgrounds, or contrast while preserving composition, identity, and visual structure.

Key features include prompt-guided editing, support for up to 4 input images, seed control for repeatable results, and flexible output options such as hosted URLs or Base64 encoding. The model excels in "update, not regenerate" workflows, making it ideal for high-volume creative tasks where speed and reliability are critical, with optimizations for maintaining geometry, materials, lighting, and overall realism.

What makes it unique is its turbo-speed performance combined with high prompt adherence, sharp details, accurate color control, and superior consistency across edits, built on the FLUX.2 [dev] architecture for photorealistic results and professional-grade editing without full regeneration.

• Architecture: FLUX.2 [dev] from Black Forest Labs
• Parameters: Not publicly specified
• Resolution: 512 to 2048 pixels width/height; presets include squarehd, square, portrait43, portrait169, landscape43, landscape169; custom width/height supported
• Input/Output formats: Input: 1-4 image URLs or Base64 data URIs; Output: PNG images as URLs or Base64; supports numimages up to unspecified limit (default 1)
• Performance metrics: Ultra-fast inference; costs scale with megapixels (e.g., $0.008 per MP input/output); guidance_scale default 2.5; seed for repeatability (-1 for random)

• Use fixed seeds for repeatable edits and testing variations; random seed (-1) for new outputs each run
• Best practices: Provide clear, descriptive prompts like "change the weather to winter" or "increase contrast while keeping the face and pose"; enable prompt expansion for better results
• Common pitfalls: Limit to 4 input images max; large Base64 inputs may slow performance
• Quality vs speed trade-offs: Turbo mode prioritizes speed with reliable edits; adjust guidance_scale (default 2.5) higher for stricter prompt adherence, potentially at minor speed cost
• Prompt engineering tips: Use natural language for targeted changes; specify hex colors for precise control; combine with multi-image references for context

• Optimal parameter settings: guidancescale 2.5-4.0 for balance; imagesize "squarehd" for high detail; enablesyncmode true for immediate results in API use
• Prompt structuring advice: Be specific and iterative, e.g., start with "make it winter" then refine to "add snow with studio gray background"
• How to achieve specific results: For identity preservation, use prompts emphasizing "keep pose and face"; multi-image inputs for reference-based consistency
• Iterative refinement strategies: Run with fixed seed, preview, tweak prompt slightly, and re-run for controlled variations
• Advanced techniques: Combine hex color prompts like "change background to #808080"; use enablebase64output for direct embedding; test numimages >1 for batch edits

• Precise image-to-image editing with natural language prompts and hex color control
• Preserves composition, subject identity, geometry, materials, and lighting during edits
• Supports up to 4 input images for multi-reference context and targeted transformations
• Generates photorealistic outputs with sharp details, accurate colors, and high realism
• High versatility for subtle retouching to bold restyles; strong prompt adherence and consistency
• Repeatable results via seed control; fast for high-volume workflows

• High-volume creative workflows like quick iterations on product visuals or character designs
• Targeted restyling such as weather changes, background swaps, or contrast adjustments in photoreal pipelines
• Professional design tasks requiring structure-preserving edits, e.g., brand-safe consistency in advertising
• Reference-driven edits for maintaining visual identity across image series in creative projects
• Subtle corrections or bold transformations in high-resolution pipelines without full regeneration

• Experimental features: Prompt expansion option improves results automatically when enabled
• Known quirks: Input images resized to 1MP for processing; only first 4 of >4 images used
• Performance considerations: Ultra-fast with costs per megapixel; sync mode blocks for immediate output
• Resource requirements: Handles large files via URLs preferred over Base64 for speed
• Consistency factors: Excellent preservation of layout and subjects; seed ensures reproducibility
• Positive user feedback themes: Praised for speed, precision, and realism in editing workflows
• Common concerns: May require iterative prompting for complex changes; NSFW detection included

• Primarily optimized for editing existing images, not pure text-to-image generation
• Resolution capped at 2048 pixels; custom sizes must fit 512-2048 range
• Potential minor inconsistencies in very complex multi-element edits without iterative refinement