Black Forest Labs
FLUX Kontext is the right pick when the cutout-and-recompose pipeline must preserve a specific identity exactly — a character's face, a model's outfit, a product's exact silhouette and label. Where Nano Banana 2 re-renders the subject inside the new scene (which can drift slightly), Flux Kontext's entire architecture is built for outfit/scene swaps while preserving identity intact. On Martini, the pipeline is: source image → background removal tool → Flux Kontext for the recompose step with identity-preservation prompts. Use this scenario specifically when working with branded talent, AI characters that must hold across many composites, or product hero shots where the SKU details cannot drift.
Pick Flux Kontext over Nano Banana 2 in three specific cases: (1) the subject is a branded spokesperson or AI character whose face/identity must hold across a series of composites; (2) the subject is a product hero shot where exact silhouette, label position, and color must match the SKU spec sheet; (3) the workflow is "swap the background only, leave the subject untouched" rather than "drop the subject into a new render." For looser compositing where some subject drift is acceptable in exchange for stronger environment integration, Nano Banana 2 is the better pick. Make this decision before building the canvas — the cutout step is the same, only the recompose model changes.
For an AI character, generate the canonical reference once on Nano Banana 2 (the strongest face-locker in the Martini lineup), then use that reference as the source for every subsequent Flux Kontext composite. For a real photographed subject, use the highest-quality source available — Flux Kontext's identity preservation depends on having clear identity signal in the source. Avoid sources where the face/product is partially occluded, motion-blurred, or shot at an extreme angle: the model can only preserve what it can see. Upload at 2K minimum.
Add a Tool node for background removal (Bria RMBG or workspace default), connect the source Image node, get a transparent PNG. Then add a Flux Kontext Image node and pin the cutout. Prompt with explicit identity-preservation language: "preserve the exact face, hair, and outfit of the subject in reference 1; place them in [new scene description]; do not modify identity." This pattern tells the model the subject is canonical and only the background should change. Flux Kontext is built for this prompt structure — it interprets identity language stricter than Nano Banana 2 and produces composites where the subject reads as identical to the source while the surrounding scene is new.
For multi-composite series (an AI influencer in 12 different outfits, a product across 8 lifestyle scenes), generate all composites and lay them out side-by-side on the canvas. Quick visual check: does every face still read as the same person? Does the product silhouette stay consistent? If one composite drifts (the model occasionally interprets the prompt loosely), regenerate that single shot with stricter identity language ("DO NOT change the subject's appearance from reference 1"). The canvas template pattern is designed for this: save the working composite arrangement once, then duplicate the template per outfit/scene/SKU and only change the prompt's scene description.
AI influencer composite — explicit identity-preservation directive plus new scene description. Flux Kontext interprets the "do not modify identity" instruction strictly and holds the face/outfit while changing only the environment.
preserve the exact face, hair, makeup, and outfit of the subject in reference 1; place them standing in a sunlit Tokyo alley with neon signs in the background; three-quarter angle; do not modify identity
Product hero shot composite — preserves SKU-critical details (silhouette, label, color, finish) while swapping the environment. Use this when the new scene is a brand-controlled studio backdrop rather than a photographed location.
preserve the exact silhouette, color, label, and surface finish of the product in reference 1; place it on a polished concrete surface with cool studio lighting from above; minimal shadow; commercial product photography style
Flux Kontext is the identity-preservation pick. For looser composites where environment integration matters more than strict identity, Nano Banana 2 produces stronger scene fit.
Use explicit identity-preservation language in prompts: "preserve the exact face/outfit/silhouette of subject in reference 1; do not modify identity." The model reads this directive stricter than implicit identity prompts.
Source image quality drives identity-preservation success. Generate the canonical reference once on Nano Banana 2 (strong face-locker), then use that reference across every Flux Kontext composite for series consistency.
For multi-composite series, lay out all variants on the canvas side-by-side and visually validate identity match before shipping. Drift on a single shot can be regenerated with stricter prompt language without redoing the whole series.
Companion tool page: `models/tools/background-removal` covers the cutout step. This how-to is the Flux-Kontext-paired identity-preservation compositing pipeline.
Flux Kontext paired with background removal is the strict-identity compositing pipeline on Martini — the right tool when the subject must hold exactly across composites and only the background changes. Trade-off vs. Nano Banana 2 paired: Flux Kontext locks identity stricter; Nano Banana 2 integrates environments more naturally but allows slight subject drift. For AI influencer content, branded character libraries, and product hero shots where SKU specs cannot vary, Flux Kontext is the safer pick. The canvas template pattern lets you build the composite arrangement once and reuse it across every outfit, scene, or SKU in the series — change only the prompt's scene description, keep the identity-preservation language constant. The full pipeline runs on Martini without external tools.
Connect FLUX Kontext with other AI models on Martini's infinite canvas. No GPU required — start free.
Get Started Free