par Black Forest LabsVidéoImage

FLUX

8 modèles · à partir de 55 crédits

Modèles de la famille FLUX

Modèle	Mode	Prix	Formats d'image
FLUX 3 Image	Texte vers image	à partir de 55 crédits	auto · 21:9 · 2:1 · 16:9 · 3:2 · 7:5	Ouvrir →
FLUX 3 Image (Edit)	Édition d'image	à partir de 55 crédits	auto · 21:9 · 2:1 · 16:9 · 3:2 · 7:5	Ouvrir →
FLUX 3 Video	Texte vers vidéo	à partir de 1 870 crédits	auto · 21:9 · 2:1 · 16:9 · 4:3 · 1:1	Ouvrir →
FLUX 3 Video (Edit)	Édition vidéo	Dépend de votre fichier	—	Ouvrir →
FLUX 3 Video (Extend)	Prolongation vidéo	Dépend de votre fichier	auto · 21:9 · 2:1 · 16:9 · 4:3 · 1:1	Ouvrir →
FLUX 3 Video (First & Last Frame)	Première et dernière image	à partir de 1 870 crédits	auto · 21:9 · 2:1 · 16:9 · 4:3 · 1:1	Ouvrir →
FLUX 3 Video (Image)	Image vers vidéo	à partir de 1 870 crédits	auto · 21:9 · 2:1 · 16:9 · 4:3 · 1:1	Ouvrir →
FLUX 3 Video (Keyframes)	Première et dernière image	à partir de 1 870 crédits	auto · 21:9 · 2:1 · 16:9 · 4:3 · 1:1	Ouvrir →

Règles de LUVI Prompt Engine pour FLUX 3 Edit Video

TARGET MODEL: FLUX 3 Edit Video (Black Forest Labs FLUX Video Edit, served on Atlas as black-forest-labs/flux-3/edit-video). Video-to-video EDITOR, not a generator: one source clip plus one written instruction. Length, framing, camera movement, cuts, timing, frame rate and the source's sound all come from the clip and are kept frame for frame. Faces, products and anything the prompt does not name stay as shot. Output is 24 fps, at most 720p.

Règles à respecter

DESCRIBE THE CHANGE, NOT THE SHOT. Everything the prompt does not mention stays as filmed, so the shortest prompt that names the change is usually right. Vendor examples, verbatim: "Remove the orange bucket." · "Replace the mackerel with lobsters." · "Add a seagull standing on the corner of the crate." · "Make it snow heavily; give the scene a cold winter palette."
NEVER RE-DESCRIBE WHAT SHOULD STAY. Restating the subject, the camera, the setting or the sound invites the model to re-render them. Do not add camera moves, shot lists, HARD CUT, timecodes, durations, resolutions, aspect ratios or audio layers — none of them can be set here, and writing them asks for changes nobody wanted.
Detail goes WHERE THE EDIT IS. The more shape, material or action you write for the changed element, the more the model changes it — so be specific about the new thing and silent about the rest.
ADDITIONS say where the element goes and what it does ("a seagull standing on the corner of the crate"), and how it is lit so it sits in the scene.
REMOVALS say what should be visible where the object was, so the fill holds across frames.
Use the verbs the vendor uses: Remove · Replace X with Y · Add · Change · Make (it snow / him say …) · Restyle. Keep the user's own nouns — the model has to find the thing they mean in the frame.
DIALOGUE CHANGES: quote the new line and name who says it (the vendor's own example: Make him say 'Fresh mackerel, four for ten.') and keep it about as long as the speech it replaces — extra words beyond the time the original line occupies will not all render or lip-sync. Translation of a line is supported: name the target language.
On-screen text: quote the new copy exactly; perspective is kept.
Sound is kept unless the prompt asks to change it. A silent source returns a silent result — a prompt cannot add a soundtrack to a silent clip.
Prompt length: 1 to 4,096 characters.

Ce qui fonctionne le mieux

One idea per request. Several changes can go in one prompt, but staged edits (feed the result back as the source) keep more fine detail on bigger jobs.
Small, clearly described ACTION changes work best; large changes to what people do are less reliable — say so rather than promising them.
Restyle requests name the medium only (Atlas's example: "Adjust the video style to a comic book style."); the performance and camera movement stay as shot.
Relighting, weather, season, time of day, background or whole-location swaps are single sentences — no mask, no green screen.
There is no seed: the same instruction on the same clip can come back different. Never write a rewrite that implies a variation of a previous render.

Issu des règles de LUVI Prompt Engine pour cette famille : les mêmes que celles qu'appliquent le réécriveur de l'espace de travail et le connecteur MCP de LUVI.

Règles de LUVI Prompt Engine pour FLUX 3 Image

TARGET MODEL: FLUX 3 Image (Black Forest Labs, served on Atlas as black-forest-labs/flux-3/text-to-image and /edit). Native output up to 4K, strong layout following and clean in-image text. The control surface is the prompt plus settings: no negative_prompt, no weights, no seed.

Règles à respecter

Plain sentences, SUBJECT FIRST, then the action, setting, light, composition and style. Vendor template: "Describe the subject, setting, light, and framing."
SAY WHAT YOU WANT, NOT WHAT YOU DON'T. Atlas's own guide: "a deserted beach at dawn", not "a beach, no people" — naming an exclusion can put it in the frame.
In-image text goes in double quotes, with its style and placement: a van covered in graffiti that spells "FLUX" in chrome bubble letters. Keep every quoted string verbatim.
Exact colours as hex codes paired with a name: walls painted #0047AB (cobalt blue).
Concrete visual detail beats praise words: "soft window light from the left, 28mm lens, fine film grain" beats "stunning". Lens and grain terms are documented here (unlike FLUX 3 Video).
Aspect ratio and resolution are settings, never prose — do not write "16:9", "4K" or "2k" into the prompt.
No weight syntax, no "--ar"/"--no" flags, no "Negative prompt:" labels — none of them exist on this surface.

Ce qui fonctionne le mieux

Describe placement the way you want it composed ("a small logo in the lower-right corner, the product centred on the left third"); the model follows layout closely.
Bounding boxes ([top, left, bottom, right] on a 0–1000 grid, appended to the prompt) are BFL-documented but untested on LUVI's surface — never invent them; keep them verbatim if the user wrote them.

Issu des règles de LUVI Prompt Engine pour cette famille : les mêmes que celles qu'appliquent le réécriveur de l'espace de travail et le connecteur MCP de LUVI.

Règles de LUVI Prompt Engine pour FLUX 3 Video

TARGET MODEL: FLUX 3 Video (Black Forest Labs — native joint audio-video, preview model). One pass renders the picture AND its sound together: dialogue with lip-sync, ambience, effects and music, across 13+ languages including Turkish. The control surface is thin — no negative_prompt, no cfg/guidance, no weights, no seed, and nothing rewrites your prompt. What you write is what is read.

Règles à respecter

ONE framing term + ONE movement term + ONE subject action per sentence. Verbatim from the vendor: "low tracking shot" is clear, "low aerial handheld orbit push-in" usually is not. The discipline is on CAMERA vocabulary, not on subject vocabulary — camera and subject may both move.
Order: subject and action FIRST, then framing, movement and style around them. The camera phrase is a short prefix, never a paragraph. Do not front-load the way the FLUX image guides do — that rule is not restated for FLUX 3.
STATE THE MOTION MAGNITUDE. It is a documented slot, not an inference: slow, abrupt, weightless, chaotic, precise, cinematic, documentary. The vendor's own examples always prepend it.
AUDIO IS WRITTEN IN, in four layers, and you do not need all four — one or two is usually right, because a busy mix is a documented failure. Speech (who, the exact words, how) · Ambience (the sound of the place) · Effects (tied to a VISIBLE action: "the mug clicks against the saucer when he sets it down") · Music (style plus where it sits in the mix).
DIALOGUE IS QUOTED AND GIVEN A VISIBLE SPEAKER. A quoted line with no visible speaker and no explicit voiceover/narration cue MAY RENDER AS ON-SCREEN TEXT instead of speech — this is the single best-documented failure of the model. A visible speaker gives it a face to lip-sync; an off-screen line needs the word "voiceover" or "narration".
NEVER ASK FOR SILENCE IN PROSE. "quiet room tone" collapses into static or dead air. Name the sound you want instead. Full silence is the generate_audio parameter, not a sentence. Partial silence is fine as content: "for the final two seconds, only rain against the window".
A MIX NEEDS A FOREGROUND LAYER OR IT DRIFTS TOWARD INAUDIBLE. Measured across ten clips: every soundscape built only from ambience and effects landed between -36 and -61 LUFS — the bottom of that range is effectively silence — while every clip carrying dialogue or a music bed sat between -18 and -32. Naming the sounds is necessary but not sufficient. When the audio matters and the scene has no speech, give it a music cue and say where it sits in the mix.
Label each spoken line's language, and keep lines SHORT — a short line in a longer clip is safer than copy written to fill every second. Multi-speaker: attribute by visible ROLE, never by name, with short separate turns and no overlap.
Lens and framing are CATEGORY WORDS, not numbers. No f-stops, no film stocks, no camera-body names, no hex colours — all of that is image-guide vocabulary with zero documented effect here. "Wide angle (24mm)" is the ONLY focal length in the entire video documentation, and the number is redundant with its bucket. Depth of field goes by named technique: "shallow (sharp on subject, blurred background)". Colour goes in prose: "warm backlight with soft rim — amber, cream, walnut".
Use real grip terms, and the right ones: the vendor honours pedestal, crane/boom, trucking, dolly in, arc shot, camera roll, locked-on and steadicam follow as DISTINCT moves. "Push-in" and "pull-back" are not enum terms — write "dolly in" or "push through". Bare "zoom" is not documented — only "dolly zoom".
Concrete nouns and verbs the camera can actually see. Vague adjectives leave the result to chance; you are directing a scene, not describing a collection of objects.
Length is not the goal. Short prompts leave details to chance; over-stuffing makes MOTION LESS COHERENT. About 330 words is the vendor's own anchor for a 10-second, 4-shot sequence.
Negation is first-party practice here despite there being no negative_prompt field — but ONLY for audio layers and screen furniture: "No on-screen text, no subtitles." · "no announcer delivery" · "no music, no second voice". Never negate into silence.
Kill ad-copy voice. If a read sounds like an advert, the fix is a person plus a recording setup plus a social situation, not praise words: "close and dry, warm mid register, lightly amused".

Ce qui fonctionne le mieux

MULTI-SHOT INSIDE ONE GENERATION is this model's headline capability, and the only place continuity is model-enforced rather than prompt-hoped — there is no seed and no style reference, so prefer one multi-shot generation over stitched clips. HARD CUT is the documented cut token: "SHOT ONE: ... HARD CUT. SHOT TWO: ...". Character, look and continuity hold across the cuts with a single audio bed carrying through.
Continuity is carried by a VERBATIM-IDENTICAL SUBJECT PARAGRAPH repeated in every shot, plus one global style-and-colour block.
Beat density: two or three beats per 5 seconds. The vendor's 10-second example is 4 shots of 2.5s.
TIMECODES ARE ORDERING, NOT MARKS. Measured on this surface: a single HARD CUT written at 5.0s of a 10s clip rendered as cuts at 1.46s and 7.71s, while shot order, identity and dialogue placement all held. Write beats as a timeline when the action has to land in sequence — never promise a frame-accurate in-point.
Three registers, pick by need: the one-liner default "[camera] shot of [subject] [action] in [environment]. [supporting visual and motion details]"; labelled fields when tuning levers (Camera shot / Subject + action / Depth of field / Lighting + palette / Motion / Style); the full schema for multi-shot (Core summary · Scene · Subject description · Dynamic narrative · Audio · Style & colour).
Style is prose only. Vetted terms from the vendor's own examples: "cinematic naturalism", "gritty documentary realism", "epic western realism", "fine grain adds gritty, tactile realism". There is no raw mode — ask for realism AFFIRMATIVELY (grain, pores, unretouched texture); you cannot ask for plastic sheen to be removed.
A director's name only when paired with the mechanical description that does the work ("staged like a Wes Anderson set — uniformed staff in neat formation, deadpan symmetry"). The name alone underperforms.
NO TAKE IS REPEATABLE on this surface: there is no seed and no draft loop, and two identical submits differ completely. Never write a rewrite that implies a variation of a previous render.

Issu des règles de LUVI Prompt Engine pour cette famille : les mêmes que celles qu'appliquent le réécriveur de l'espace de travail et le connecteur MCP de LUVI.