FLUX scene specification
Convert a reference image into a clear FLUX prompt built around subject, action, style, and context. The generated result favors natural language, explicit object relationships, affirmative constraints, and enough scene detail for FLUX to follow the intended composition.
A white wristwatch centered on a cool gray studio surface, viewed directly from above, with the strap extending diagonally toward the upper right and a soft contact shadow beneath the metal case.
Subject
white wristwatch
Position
centered, top-down
Material
metal and matte silicone
Constraint
clean surface, sharp product
FLUX scene input
Generate the six outputs, then inspect the FLUX version for explicit positions, object relationships, and positive descriptions instead of a negative-prompt list.
No image selected
Add an image to create all six prompt formats.General · Midjourney · FLUX · Stable Diffusion · ChatGPT Image · Nano Banana
How FLUX reads a scene
The instruction tells the model what is present, what each subject is doing, how the image should look, and where the scene takes place. Black Forest Labs describes a useful starting structure as subject, action, style, and context. The format is flexible. Natural language works well, and a complete sentence often carries spatial relationships more clearly than a stack of disconnected tags.
This page turns visible evidence into that kind of instruction. It reads the main subject, environment, composition, lighting, palette, camera feel, materials, style, and mood. The analysis then writes a FLUX-specific version rather than copying the General prompt. That version can say which object is left or right, what overlaps, where a person faces, and how the light interacts with a surface.
The generated description is an inference. It is not the original production record. The pixels cannot disclose a seed, hidden prompt, model variant, reference stack, post-processing step, or exact camera setting. The useful task is to rebuild the scene specification from what can be seen, then let you edit uncertain details before generation.
FLUX model variants do not all expose identical controls. Current Black Forest Labs guidance covers FLUX.1, FLUX.1 Kontext, and FLUX.2, with model-specific differences called out in its documentation. Keep the descriptive prompt portable, then set resolution, aspect ratio, seed, guidance, or reference inputs inside the destination you actually use.
Spatial relationship map
A reference image is full of relative facts: one object is behind another, a face turns toward the window, a small figure establishes scale. Those facts deserve more space than generic quality words.
Open the generated text with the concrete subject and its defining visible attributes. Use age group rather than guessing identity, describe clothing instead of personality, and name the product material instead of calling it premium. The first sentence should let a reader sketch the main silhouette.
Tell the FLUX prompt what the subject is doing and where it faces. “Courier viewed from behind, standing beside a motorcycle” prevents several plausible but incorrect arrangements. For still life work, action becomes orientation: a watch lies flat, a strap crosses another strap, or a glass sits partly outside the frame.
Write left, right, foreground, background, above, below, behind, and near when those relations define the image. A FLUX prompt gains control from explicit placement. Avoid impossible camera statements or relationships that conflict with the reference perspective.
Name the medium, camera treatment, and lighting setup in observable terms. Large diffused source, hard afternoon sun, volumetric street haze, translucent watercolor, and shallow depth of field give a FLUX prompt usable rendering cues. Connect every mood label to those visible choices.
For FLUX.2, official guidance recommends describing what you want because the model does not support a separate negative prompt. Ask for an empty street, clean typography, sharp focus, or an uncluttered surface. The affirmative FLUX prompt describes the desired state instead of ending with a long ban list.
Positive constraints
A common Stable Diffusion habit is to paste a large negative prompt below every generation. That habit does not transfer cleanly. Black Forest Labs states that FLUX.2 has no negative prompts and advises users to describe what they want. “Sharp focus throughout” is clearer than “no blur,” and “an empty platform” gives the scene a positive state instead of asking the model to reason through an exclusion.
The same principle helps a FLUX prompt generated from a reference. If the image contains no text, write “clean unbranded product photograph” when branding absence matters. If the background is sparse, describe the few visible elements and the amount of open space. Avoid adding ten “no” clauses that give unwanted concepts extra attention.
Precise color and typography need object-level attachment. Current FLUX guidance recommends associating a color or hex value with a specific object. A FLUX prompt should say that the car is #D94832, not merely list a hex code at the end. For text inside an image, provide the exact wording, location, hierarchy, and design context, then expect to inspect spelling after generation.
Complex work can use structured prompts or multiple references in supporting FLUX products. Keep those controls outside the prose unless the destination expects them. This generator returns readable natural language because you can audit it quickly, change a relationship, and paste it into different FLUX interfaces without carrying application-specific JSON by accident.
Relation-first revision
The generated text becomes easier to improve when you treat it like a shot description. Verify geometry first, rendering second, and decorative detail last.
Before uploading, decide what makes the reference useful. It may be the top-down symmetry, the distance between two people, the scale of a portal against a figure, or the way a product catches light. Make that relationship the first test for the FLUX prompt, because a FLUX prompt is only useful when it preserves the scene logic you care about.
Read the FLUX prompt once without judging style. Check that every important object exists and occupies the correct part of the scene. Remove invented props. Correct ambiguous pronouns. If two objects share a color, name each one rather than asking the model to infer which description belongs where.
Use the edited FLUX prompt with the required canvas and a suitable model. Do not add several references, automatic upsampling, and experimental guidance values at once. A simple baseline reveals whether the written scene is coherent before other controls reshape it.
If the subject is correct but the layout drifts, rewrite placement. If geometry holds but the surface feels plastic, strengthen material and light behavior. Keep one FLUX prompt variable under review at a time, then save the version that produced a repeatable improvement.
Scene failures
Object lists are easy. Relationships carry the composition. Most repairs start by replacing a loose list with a sentence that explains where things belong.
A FLUX prompt made of “cinematic, 8k, masterpiece, detailed” spends words on praise and leaves the scene undecided. Replace those labels with subject, action, camera position, light source, surface behavior, and the relationship that defines the reference.
Top-down, eye-level, and low-angle cannot all describe the same view. A person cannot face the camera and be shown strictly from behind. Remove contradictions from the FLUX prompt before adding detail, because the model must resolve them somehow.
A long exclusion list is not a substitute for a desired scene. Rewrite the FLUX prompt as “sharp product, clean blank face, empty neutral surface” and check the current capabilities of the chosen FLUX model or interface.
The tool may infer a lens feel or lighting setup from visual evidence, but it cannot read hidden metadata from a flattened image. Treat those details as generation directions. Do not present a generated FLUX prompt as proof of how the source was captured or created.
FLUX syntax questions
A FLUX prompt is the text description used to direct a FLUX image model. A practical version defines the subject, action, style, and context, then adds placement, lighting, materials, camera treatment, color, and constraints where they affect the result.
Yes. Official Black Forest Labs guidance supports natural language and says there is no single correct format. Complete sentences often help when the image depends on spatial relationships, multiple subjects, exact text placement, or a sequence of visual conditions.
Black Forest Labs states that FLUX.2 does not support negative prompts. Describe the desired state instead. Other versions and third-party interfaces can differ, so check the documentation for the exact model and application you use.
No. An image cannot reveal the exact FLUX prompt, seed, model variant, reference images, guidance, or post-processing history. The service writes a new instruction from visible evidence and lets you edit uncertain details.
Use camera language when it communicates a visible photographic treatment. It can guide perspective, depth of field, texture, and contrast. Treat inferred settings as creative simulation, not verified EXIF information from the source image.
One uploaded image uses one credit and generates six versions, including the FLUX prompt. Visitors receive three daily credits. Signed-in users can save and edit History, while batch input is limited to Pro subscribers.
Write the next scene specification
Upload one image, verify the subject and object relationships, then edit the FLUX prompt until the scene can be understood without seeing the reference. Describe the state you want, attach color and material to the correct objects, and move model controls into the FLUX interface where they belong. A strong FLUX prompt maps geometry before it decorates the scene. A reusable FLUX prompt also keeps application controls separate from the visual brief. The final FLUX prompt should read like one coherent shot description.