Verified against Flux · 2026-07-26
Keep a character consistent across multiple scenes with Flux.1 Kontext
A sequential in-context editing workflow for Flux.1 Kontext that locks a character's face, outfit, and proportions across a series of new scenes by editing forward from one reference image each time, instead of re-describing the character from scratch and getting a visibly different person every generation.
The prompt
Ready to copy — highlighted parts are example details you can swap.
REFERENCE IMAGE Attach reference_v1.png — a woman with a shoulder-length auburn bob, freckles, wearing a forest-green field jacket, generated in the first pass of this series as the input image for this edit. This is the character whose identity must carry forward — an identity reference, not a style reference. WHAT MUST STAY IDENTICAL Keep this exact character's face, the freckle pattern across the nose and cheeks, and the exact shoulder-length auburn bob with a center part, and the forest-green field jacket with brass buttons, worn open over a cream shirt unchanged from the reference image. Do not redesign, restyle, or "improve" any of these — treat them as fixed facts about a real person, not a starting point to riff on. WHAT CHANGES IN THIS EDIT now crouched examining a trailhead map at the edge of a pine forest, three-quarter angle from camera-left, mid-morning HOW THE CHANGE SHOULD READ The character should look like the same photograph's subject placed into a new moment, not a new illustration of a similar-looking person — same skin tone rendering, same approximate age, same build, same hairstyle, carried forward exactly, with only the new scene and pose actually changing. LIGHTING CONTINUITY switch from the flat overcast light of the reference to dappled morning sunlight coming through pine branches. If the new scene's lighting genuinely differs from the reference — a different time of day, indoors versus outdoors — let the light on the character update to match the new environment, but the change should look like real light hitting the same person, not a different person rendered under different light. SEQUENCE NOTE This is edit 4 in a sequence of scenes using the same character. Treat the reference image as ground truth for identity even if a prior edit in the sequence drifted slightly — correct back toward the original reference rather than compounding the drift forward. FRAMING AND CROP FOR THIS SCENE State whether this edit should keep roughly the same crop and camera distance as the reference (a close portrait staying close, a wide environmental shot staying wide) or deliberately change it for this specific new scene — Kontext will otherwise sometimes default to whichever framing was most recently generated in the sequence rather than the framing this particular new scene actually calls for, so name the intended shot type explicitly rather than leaving it implicit. WHAT TO DO IF IDENTITY DRIFT IS VISIBLE If, on review, the face, hairstyle, or outfit in this output has drifted noticeably from reference_v1.png — a woman with a shoulder-length auburn bob, freckles, wearing a forest-green field jacket, generated in the first pass of this series, do not accept it and build on top of the drifted version — regenerate this specific edit directly from reference_v1.png — a woman with a shoulder-length auburn bob, freckles, wearing a forest-green field jacket, generated in the first pass of this series again with the same new_scene_and_pose instruction, rather than letting an imperfect result become the new baseline for the next edit in the series. OUTPUT One image: the same character from reference_v1.png — a woman with a shoulder-length auburn bob, freckles, wearing a forest-green field jacket, generated in the first pass of this series, now in the scene and pose described above, with no visible seam, warping, or identity drift around the face compared to the reference.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Kontext is architecturally different from base Flux.2 text-to-image generation: it takes an actual image as input and edits it in-context rather than generating fresh from a text description alone. That means the prompt's real job is to specify what to preserve and what to change relative to that input, not to redescribe the whole scene from zero — which is exactly why WHAT MUST STAY IDENTICAL exists as its own section, distinct from WHAT CHANGES IN THIS EDIT, rather than one blended description. Naming identity_features and outfit_description explicitly, instead of just saying "the same character," matters because Kontext — like any image model performing an edit — has a real tendency toward gradual identity drift over a sequence of edits-of-edits: a slightly different nose, a slightly lighter hair color creeping in generation over generation. Being explicit about the exact traits gives the model concrete anchors to check the new frame against, rather than a vague continuity it can only approximate. The sequence note — correct back toward the original reference rather than compounding drift forward — targets a specific failure mode of chained edits directly: if edit 3 already drifted slightly from edit 1, and each new edit is generated from the most recent output rather than the original, small errors compound generation over generation until the character is recognizably different by edit 6 or 7. Re-anchoring explicitly to the original reference image every time, not the previous output, is the practical fix for that compounding. The lighting-continuity instruction matters because Kontext can fail in either of two directions without it: over-preserving the original scene's flat lighting even when the new environment obviously wouldn't have it, or under-preserving in the other direction by letting the character's actual skin-tone rendering shift along with the lighting change. Separating "let the light update" from "keep the same person under it" as two distinct instructions is what prevents both failure directions at once.
What you get back
A sequence of four images, all recognizably the same auburn-bob character in the forest-green field jacket — one on a trailhead, one crossing a stream, one at a summit viewpoint, one setting up a tent — with no visible drift in face shape, freckle pattern, or jacket detail between them.
Verified against
Flux Flux.1 Kontext [pro] · 2026-07-26
Changelog
- 2026-07-26 — Initial publish, verified against Flux.1 Kontext [pro] across a four-image sequential-edit test.
Need this built into your business?
If a prompt isn't enough — what Scult builds, built and maintained for you — that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
