Black Forest Labs launched Flux 3 Image on October 1, 2026, completing the image arm of its multimodal Flux 3 family with a system that accepts structured JSON scene layouts and preserves every pixel outside a selected region with bit-identical numerical fidelity — a capability gap that has made AI image editing unreliable for production workflows and LLM-driven pipelines.

The release shifts the paradigm that has governed AI image editing since the first inpainting tools: rather than redrawing an entire frame and hoping surrounding regions look acceptably similar, Flux 3 Image constrains generation to named bounding boxes while leaving outside pixels mathematically unchanged. For developers building agentic content pipelines or iterative design tools, that distinction is the difference between a tool that accumulates visual drift across editing sessions and one that can be treated as reliable, deterministic state.

Black Forest Labs has followed a consistent pattern with every prior generation of its Flux models: a hosted API release precedes open-weight access by weeks to months. An open-weight version of Flux 3 Image is expected within weeks, with no specific calendar date announced.

What Is a Bounding-Box Layout, and Why Does It Matter?

The most technically significant feature of Flux 3 Image is its bounding-box composition system. Rather than inferring object placement from prose description alone — a method that yields inconsistent spatial results across generations — the model accepts a structured JSON array of scene elements in which each element carries a unique identifier, a natural-language description, and a set of coordinates.

Those coordinates map to a normalized 0-to-1000 grid measured from the top-left corner of the frame, in the order [top, left, bottom, right]. The grid is decoupled from pixel resolution and aspect ratio: a layout specifying [150, 50, 850, 650] for a subject places that subject in the same relative position whether the output is a 768-pixel square or a native 4K frame at 5,456 × 3,072 pixels (approximately 16.8 megapixels). This resolution independence is architecturally significant for applications: the element table can be authored once, stored as project state, and applied consistently across output sizes without recalculation.

A layout for a product composite, for example, might define background_1 as a full-frame environment, product_1 with coordinates placing it in the left third, and text_1 in the upper right — each with its own description, each independently editable in subsequent turns. Developers can author these element tables directly in code, or they can instruct a language model to generate the table from a brief natural-language description and pass the output directly to the Flux 3 Image API.

How the Editing Pipeline Works Without Drift

The second headline capability is what BFL describes as bit-identical pixel preservation in multi-step editing. When a user or application selects a bounding box and issues an edit instruction — recolor this element, replace this object, move this element to new coordinates — the model regenerates only the specified region while leaving all pixels outside that box numerically unchanged.

BFL’s own published pixel-difference measurements confirm the claim: a cat-removal operation left 80.7 percent of the total image pixels identical; a bird-addition operation left 86.4 percent unchanged; a shirt-print operation left 89.7 percent intact. A hair-recolor operation, which necessarily bleeds color into immediately adjacent regions, left 67.8 percent unchanged — the lowest figure in BFL’s published examples, and a meaningful disclosure about the limits of the guarantee.

This differs fundamentally from prior inpainting approaches, where the model re-renders the masked region by sampling from the surrounding context — a process that introduces subtle pixel drift in neighboring areas even when only the masked area is supposed to change. In a pipeline with many successive edits, that drift compounds. Flux 3 Image’s architecture avoids it by treating the element table and its stable IDs as persistent project state: re-issuing an element’s ID with a new description or new target coordinates regenerates only that element, with everything else locked.

The Agentic Workflow Implication

Because element IDs and coordinates are stable JSON state, an LLM agent can treat the element table as a persistent project file across conversation turns. The agent can generate an initial composition, read back the element IDs, issue targeted edits to individual elements in subsequent API calls, and build a complex composite iteratively — without any regeneration of unedited regions and without any background drift accumulating between turns.

This is architecturally different from how prior image-editing APIs work. Tools that require regenerating from scratch on each edit, or that accept only natural-language masks, give agents no reliable way to address a specific element without risking the rest of the frame. The bounding-box system’s stable IDs function as object handles — a concept familiar from code and database systems but previously unavailable in image generation. For developers building automated creative pipelines, this is as significant as the spatial precision.

The system also includes a prompt upsampler — a component that expands a user’s short instruction into the detailed caption format the model was trained on. Box IDs and coordinate arrays bypass this rewrite step and reach the model unchanged, which limits the scope of any generation to the specified region exactly as the developer intended.

Ten Reference Images and Native 4K

Flux 3 Image accepts up to ten reference images per generation, supplied as URLs or base64-encoded data and cited inline in the prompt using tokens assigned in upload order: ref_image_0 through ref_image_9. The model determines each reference image’s placement and scale before rendering the complete frame — a pipeline suited to product composites and lookbooks assembled from existing visual assets.

Native output reaches approximately 5,456 × 3,072 pixels without a separate upscaling pass. BFL demonstrates this capability with a soba shop scene in which hand-painted Japanese characters roughly 225 pixels tall remain legible at native resolution — a result the company attributes to generating at full resolution rather than enlarging a smaller output after the fact. Upscaling typically softens fine textures and small text; native generation avoids this by computing full detail from the start.

How Flux 3 Compares to Flux 2 and to Ideogram 4.5

Flux 3 Image slots into a four-part multimodal architecture that BFL began disclosing in July 2026: Flux 3 Video (public since August 4, 2026), Flux 3 Image (API and playground access as of October 1, 2026), Flux 3 Action (early access, purpose not fully disclosed), and the forthcoming Flux 3 Dev open-source release. BFL describes the family as sharing one multimodal model trained jointly across image, video, and audio data using its Self-Flow architecture — a self-supervised flow matching framework that jointly learns generation and understanding across modalities within a single set of weights.

Compared with Flux 2 — which shipped in November 2025 as a 32-billion-parameter dense transformer limited to image generation — Flux 3 Image’s key advances are the structured bounding-box layout system (absent from Flux 2), the bit-identical pixel preservation guarantee (an improvement over Flux 2’s inpainting approach), and native 4K output (versus Flux 2’s 4-megapixel ceiling). Flux 2’s standing had slipped in Artificial Analysis text-to-image evaluations by August 2026, providing context for why BFL needed a differentiated next generation.

One day before Flux 3 Image shipped, Ideogram released Ideogram 4.5 Precise Edit — a model that applies natural-language instructions to a source image and restores unchanged pixels from the source so surrounding detail stays intact across repeated edits. The simultaneous emergence of pixel-preservation editing from two competing labs confirms that the capability is becoming a competitive standard rather than a differentiating feature of any single model.

The two approaches share the objective — lock unchanged pixels — but differ in architecture. Flux 3 Image’s system is built on explicit element IDs and coordinate arrays that persist across editing turns; Ideogram 4.5 Precise Edit operates from a source image plus a natural-language instruction and a mask, restoring unchanged pixels from the original frame. For developers who need programmatic, agent-accessible multi-turn composition, Flux 3’s JSON element table provides more deterministic control; for designers who prefer natural-language instructions with no layout-authoring overhead, Ideogram 4.5 Precise Edit may be more accessible.

Access, Pricing, and the Open-Weight Roadmap

Flux 3 Image is available through the BFL Playground and API as of October 1, 2026. BFL is offering a 50 percent introductory API discount through October 8, 2026 — meaning the current entry price of $0.0205 per image at 768px will roughly double after the discount window closes. Companies can also license commercial weights for fine-tuning and self-hosted deployment on their own infrastructure.

An open-weight version of Flux 3 Image is expected within weeks, with no specific calendar date announced. The separate Flux 3 Dev open-source product line remains on a distinct roadmap. BFL has followed this pattern with every prior generation: a hosted API release precedes open-weight access by weeks to months, and both eventually become available.

Black Forest Labs is headquartered in Freiburg, Germany, and has approximately 100 employees as of August 2026, with 30 in San Francisco. The company was founded in August 2024 by Robin Rombach, Andreas Blattmann, Patrick Esser, and Dominik Lorenz — the principal researchers behind Latent Diffusion Models and Stable Diffusion. BFL reached a valuation of $3.25 billion following a $300 million Series B in December 2025, led by AMP and Salesforce Ventures, with total funding above $450 million and investors including Andreessen Horowitz, General Catalyst, NVIDIA, Canva, and Figma Ventures.

Frequently Asked Questions

What does “bit-identical” pixel preservation actually mean, and why does it matter for agentic workflows?

Bit-identical means the numeric color values of pixels outside the selected bounding box are mathematically unchanged — not just visually similar, but the exact same numbers. This matters for LLM-driven pipelines because an agent can issue multiple successive edit instructions to different bounding boxes across multiple API calls, and the composition outside each edited region will not accumulate any visual drift between turns. In conventional inpainting, re-rendering the masked region slightly perturbs neighboring pixels, which compounds across many edits. Bit-identical preservation removes that accumulation, letting developers treat the element table as reliable project state the way they would treat a database row. BFL’s published pixel-difference measurement examples illustrate how much of each image remains untouched across a range of edit types.

How is the bounding-box system different from Ideogram 4.5 Precise Edit?

Both approaches aim to preserve pixels outside the edited region, but they architect the control differently. Flux 3 Image requires developers to author an element table — a JSON array of named IDs, coordinate arrays, and descriptions — that persists across editing turns and provides deterministic object handles. Ideogram 4.5 Precise Edit takes a source image plus a natural-language instruction and an optional mask, then restores unchanged pixels from the source after the edit. Flux 3 Image’s system gives developers and agents more programmatic control and is better suited for multi-step automated pipelines; Ideogram 4.5’s approach requires less authoring overhead for single-step design revisions. For the bounding-box editing mechanics in depth, see BFL’s bounding-box editing documentation.

What is the Self-Flow architecture, and how does it differ from how Flux 2 worked?

Self-Flow is Black Forest Labs’ self-supervised flow matching framework that jointly trains generation and understanding across image, video, and audio modalities within a single architecture. Flux 2 was trained exclusively on image data as a 32-billion-parameter dense transformer. Self-Flow treats image generation, video generation, and audio generation as instances of the same underlying learning objective, so representations learned from video inform image generation and vice versa. In practice, this gives Flux 3 Image stronger physical plausibility and cross-reference consistency compared with image-only models. The bounding-box system sits on top of this shared architecture. More technical detail on the Flux 3 Image multimodal architecture is available in BFL’s documentation.

Is the 50 percent pricing discount the launch price, or will it be permanent?

The 50 percent introductory discount is explicitly time-limited. BFL is offering it through October 8, 2026, after which pricing will roughly double. The entry-level price of $0.0205 per image at 768px will increase to approximately $0.048 per 1K-resolution image after the discount window closes. Pricing scales with output resolution. For current rates across all resolution tiers, see fal.ai’s Flux 3 Image pricing. An open-weight version of Flux 3 Image is expected within weeks for teams that prefer to self-host.