A beginner's guide to the Fibo-Edit model by Bria on Replicate

Bria's Fibo-Edit model, available on Replicate, uses JSON-based control for deterministic, pixel-precise image editing, replacing ambiguous text prompts. The 8-billion-parameter model supports tasks like recoloring, relighting, and object removal, and is trained on fully licensed data for commercial safety.

This is a simplified guide to an AI model called Fibo-Edit maintained by Bria. If you like these kinds of analysis, you should join AImodels.fyi or follow us on Twitter. fibo-edit is a structured image editing model developed by bria https://aimodels.fyi/creators/replicate/bria?utm source=devto&utm medium=referral that replaces ambiguous text prompts with JSON-based control for deterministic, pixel-precise edits. The 8-billion-parameter model operates on the Visual GenAI Language VGL paradigm, accepting a source image, optional mask, and structured JSON instruction to perform edits like recoloring, relighting, restyling, and object removal without prompt drift. Built on Fibo's foundation and trained exclusively on fully licensed data, it prioritizes transparency and commercial safety while delivering production-grade image editing with explicit, interpretable results. The key insight before using it: unlike traditional diffusion-based editors that interpret loose language descriptions, this model treats editing as a deterministic engineering task where visual parameters lighting, composition, style, color are specified explicitly in JSON format, ensuring outputs match your exact specifications rather than probabilistically drifting from intent. Professional recoloring and material editing workflows. When you need to change object colors—recolor a car from red to blue, alter fabric tones, or modify surface finishes—the structured JSON control ensures the edit affects only the target property without unintended color shifts elsewhere in the image. This is ideal for product photography, automotive visualization, and e-commerce where color accuracy and consistency across batches matter. The mask-based precision allows you to target specific regions while leaving surrounding areas untouched. Relighting and mood adjustment for photography. The model excels at adjusting lighting conditions—adding sunset warmth, creating cool shadows, simulating different times of day, or shifting overall mood. This is valuable for photo retouching, architectural visualization, and creative direction where you want to preserve composition and objects but fundamentally change the lighting atmosphere. The JSON structure lets you specify parameters like warmth, intensity, and direction without ambiguity. Consistent style and texture transformations. Converting realistic images to artistic styles oil painting, watercolor, vintage film, impressionist or applying consistent texture treatments works well with this model, especially when you're processing batches that need uniform stylistic treatment. Fashion designers, interior visualization teams, and content creators benefit from the repeatability that structured prompts provide—the same JSON edit applied to different images produces consistent style shifts. Masked region editing for targeted object manipulation. When you have a mask defining specific regions sky, background, subject , you can apply edits only within those boundaries. This enables object removal, background replacement without touching the foreground, selective detail enhancement, and region-specific color grading. The combination of image, mask, and structured instruction gives you pixel-perfect control impossible with unmasked generation. Domain-specific LoRA finetuning for specialized editing tasks. The model supports LoRA finetuning on paired input/output image datasets with JSON captions, allowing you to adapt the base model to specialized domains—real estate photography touch-ups, medical image adjustments, or specific product photography styles. The lightweight 8B architecture makes finetuning feasible without prohibitive compute requirements. The model requires natural language instructions to be converted to structured JSON through a Vision Language Model VLM —either via Gemini API default or a local model. This adds a dependency: if you use API mode, you need a Gemini API key and internet connectivity; if you use local VLM mode, mask-based editing is unsupported, forcing a tradeoff between local execution and full feature access. The structured JSON generation step also introduces latency before actual image editing begins. Input resolution is not explicitly specified in the documentation, but the model is designed for standard image dimensions typical of diffusion models. Very high-resolution images or extreme aspect ratios may encounter undocumented constraints. The guidance scale parameter is capped at 3-5 minimum 3, maximum 5 according to the schema , which is narrower than many diffusion models, potentially limiting your ability to adjust adherence to the edit instruction beyond this range. The model cannot handle truly ambiguous or creative interpretations of edits that require understanding complex user intent—it requires explicit structural parameters. If your editing task is vague "make it better," "improve the image" , the VLM-to-JSON step must infer intent, introducing failure modes where interpretation diverges from expectation. Compared to more general-purpose image generation models, this is highly specialized for editing existing images rather than generating from scratch. The non-commercial license restricts free use to non-commercial projects; commercial applications require contacting Bria directly. The open-source code is available but with this licensing constraint. The model is production-ready but relatively new released January 2026 , so long-term stability and community adoption are still unproven compared to more established editing approaches. vs. fibo: Fibo is the base generation model that fibo-edit extends for editing workflows. Choose Fibo if you're generating images from scratch with structured control; choose fibo-edit if you're modifying existing images. Fibo uses the same VGL paradigm but is optimized for generation rather than preservation of source content, so it lacks the masking and edit-specific architecture that fibo-edit provides. vs. fibo-edit/edit: This appears to be the same model or a variant hosted on Fal rather than Replicate. Both use JSON + Mask + Image for high-quality, controllable editing. If you're already on Fal's infrastructure, their hosted version may offer better integration; if you're on Replicate, use the Replicate version for consistency and unified billing. vs. fibo-edit/restyle: Restyle is a specialized variant optimized for artistic style transfer using curated, production-grade style mappings rather than free-form JSON prompts. Choose Restyle if you want reliable, predefined style presets e.g., "oil painting," "noir film" ; choose fibo-edit if you need granular control over specific visual parameters or custom edits not covered by preset styles. vs. fibo-edit/colorize: Colorize is a specialized variant focused exclusively on color-based transformations using predefined style-based commands. Choose Colorize if your primary task is color grading or tone mapping; choose fibo-edit if you need to combine color changes with lighting, style, composition, or masked region edits in a single workflow. fibo-edit is an 8-billion-parameter diffusion-based editing model built on the Visual GenAI Language VGL paradigm. It operates natively on structured JSON input rather than free-form text, enabling deterministic edits. The model supports both masked and unmasked editing pathways: masked editing allows pixel-precise regional control, while unmasked editing applies global edits to the entire image. The pipeline integrates a Vision Language Model VLM for converting natural language instructions to structured JSON. By default, it uses Gemini 2.5 Flash via API, but a local VLM briaai/FIBO-edit-prompt-to-JSON is available via the diffusers ModularPipelineBlocks for offline operation—with the caveat that local mode does not support masked editing. The model is available through diffusers library integration, Hugging Face model hub, and multiple API endpoints Bria.ai, Fal.ai, Replicate . Key technical parameters: The model prioritizes transparency and legal safety, ranking in top benchmarks for prompt adherence and output quality. It does not use ambiguous natural language descriptions, instead treating edits as explicit parameter specifications, eliminating prompt drift common to loose-description models. structured instruction is not provided. instruction for precise, deterministic edits without VLM interpretation. python import replicate Define the edit instruction instruction = "change the car color to green" Call the model via Replicate API output = replicate.run "bria/fibo-edit", input={ "instruction": instruction, "image": "https://example.com/source image.jpg", "guidance scale": 4, "seed": 42 } The output is a URL to the edited image print output Output example: https://example.com/edited image.jpg For masked editing: python import replicate output = replicate.run "bria/fibo-edit", input={ "instruction": "make the sky blue", "image": "https://example.com/photo.jpg", "mask": "https://example.com/sky mask.jpg", "guidance scale": 4, "seed": 42 } print output For advanced programmatic control with structured JSON: python import replicate import json Define structured edit as JSON string structured edit = json.dumps { "lighting": { "warmth": "warm", "intensity": "bright", "direction": "sunset" }, "mood": "golden hour", "edit instruction": "Add sunset lighting" } output = replicate.run "bria/fibo-edit", input={ "structured instruction": structured edit, "image": "https://example.com/landscape.jpg", "guidance scale": 4 } print output Q: Do I need a Gemini API key to use this model on Replicate? A: When using the instruction parameter natural language , the Replicate deployment likely handles VLM conversion server-side, so you do not need to provide a Gemini key. However, if you run the model locally with API mode enabled, you must set the GEMINI API KEY environment variable. Using structured instruction with pre-formatted JSON bypasses the VLM entirely and requires no API key. Q: Can I use this model for commercial projects? A: The open-source weights are licensed under CC-BY-NC-4.0 non-commercial only . For commercial use, you must contact Bria directly through their website. The Replicate endpoint may have different licensing terms—check Replicate's terms or contact Bria for clarification on commercial API access. Q: What is the difference between instruction and structured instruction inputs? A: instruction is a natural language string e.g., "make it vintage" that the model converts to structured JSON via a VLM, introducing an interpretation step that may not always match intent. structured instruction expects pre-formatted JSON with explicit visual parameters lighting, color, style, composition for deterministic, unambiguous edits. Use structured instruction when you need reproducible, precise control; use instruction for convenience when exact parameter specification is not critical. Q: Does this model work with masks, and if so, how do I create one? A: Yes, the model supports mask-based editing when you provide the mask input alongside image . Masks are binary or grayscale images where white regions 255 indicate areas to edit and black regions 0 indicate areas to preserve. You can create masks using image editing tools, Python libraries like PIL or OpenCV, or manual annotation tools. API mode via Gemini VLM supports masked editing; local VLM mode does not. Q: What seed values should I use, and does setting a seed guarantee identical results? A: The seed parameter accepts any integer and controls the random number generation during diffusion inference. Setting the same seed with identical inputs image, instruction, guidance scale should produce identical or near-identical results, useful for reproducible batch processing or A/B testing. If results differ slightly despite identical seeds, this may reflect minor non-determinism in GPU operations or library versions. Q: How does this compare to other image editing models like Photoshop or general diffusion models? A: Unlike Photoshop, this model automates edits via AI rather than manual pixel-level brushwork, making it faster for large batches but less precise for fine details. Compared to general text-to-image models e.g., Stable Diffusion with inpainting , fibo-edit preserves the source image structure better, applies edits more predictably due to structured JSON control, and avoids "prompt drift" where the output diverges from instruction over time. However, it is specialized for editing existing images, not generating from scratch. Q: Can I finetune this model on my own data? A: Yes, the model supports LoRA finetuning on custom datasets. Prepare paired input/output images with a metadata.csv file containing JSON captions with fields like edit instruction , mood , lighting , etc. The finetuned model adapts to domain-specific editing tasks without retraining from scratch, making it feasible for specialized workflows like real estate photography or product retouching. Q: What resolution images does this model handle, and are there size limits? A: The README and schema do not specify maximum resolution or aspect ratio limits explicitly. Based on typical diffusion architecture, it likely handles standard image sizes e.g., up to 2048×2048 or similar , but very high resolutions or extreme aspect ratios may encounter undocumented constraints. Test with your target resolution to confirm compatibility, or refer to Bria's documentation for official specifications.