Here’s a frustration I hear constantly, in different words: “I wrote the negative prompt exactly like you said, and the model ignored it.”
Sometimes the negatives really were weak. But often the prompt work was fine, and something else in the settings panel overruled it — the guidance-style dial sitting at a value that made the model a zealot for the positive prompt and a stranger to everything else. Most generators expose some form of these dials. They rarely explain how the dials fight each other. And the fight is exactly where creators lose shots without knowing why.
This article explains the three forces — guidance, strength, and negative prompts — what each one truly does, how they collide, and the tuning order I use in production so my negatives get heard instead of shouted over. One plain caveat up front: dial names and ranges differ between tools and change between versions. I’m deliberately describing behaviour, not button labels. Learn the behaviour and you can find it in any interface.
Meet The Three Dials
Think of a generation as a negotiation between three voices. The positive prompt says what should exist. The negative prompt says what should not. The guidance dial — it travels under various names, some variation of guidance, prompt strength, adherence, or CFG in tools that borrow image-model language — decides how literally the model obeys the text as a whole. Turn it up and the model clings to the words, all the words, harder. Turn it down and it drifts toward its own instincts: more natural, sometimes, and more random, often.
The third force is strength in its looser sense. In image-to-video work this is often how strongly the generation may depart from the starting frame, and in various tools it appears as motion intensity, creativity-versus-fidelity sliders, or similar couplings. Names vary wildly; the behaviour family doesn’t. One pole says stay faithful to what you were given. The other says improvise. Where you set it decides how much of your careful frame survives and how much motion energy the clip carries.
My point in grouping them: creators treat these controls as independent volume knobs. They’re not. They’re one system, and moving any one of them changes what the other two can accomplish. Which is why tuning them in isolation, by vibes, produces footage that fails in ways that feel mysterious. Sound familiar?
Negative prompts live inside this system in a specific seat: they’re part of the text, so guidance scales them too. That single fact explains most of the collisions below. Raise guidance and you amplify your negatives along with your positives — helpful until it isn’t. Raise the freedom or motion side high enough and the model may wander so far from frame and text that no negative list can corral it. The dials aren’t enemies. They’re just all holding the same steering wheel.
What Guidance Actually Amplifies
Guidance amplifies obedience, and obedience is neutral. At moderate settings, that’s what you want: the subject you described is present, the action happens, the camera intention reads. Push guidance high and the model becomes a literalist — it’ll force every noun in your prompt into the frame whether they fit together or not, exaggerate described features, and strain to satisfy conflicting instructions simultaneously.
Faces get intense. Colours get loud. Compositions get stiff, like the model is afraid to leave anything out. Creators often describe high-guidance output as “overcooked,” and the word is apt: the ingredients are all yours, cooked too hard.
For negatives, high guidance cuts both ways. Your bans are obeyed more forcefully — good — but so is everything else, and force itself becomes an artifact source. A model straining to satisfy a long prompt plus a long negative list plus fierce guidance will produce tension somewhere. And tension usually appears as stiffness, over-sharpening, or that glossy literal look where nothing breathes.
This is one half of the classic dial fight: the creator raises guidance because a prompt element was ignored, and the higher guidance makes the footage look worse in a new way that gets blamed on the negatives. Ouch.
Low guidance has its own honest uses. It lets the model smooth over awkward phrasing, find natural motion, and behave more photographically. Some of my favourite ambient shots — weather, water, empty streets — run on the gentler side, where the model is essentially art-directing and my prompt is a suggestion. But low guidance on a precision job is abdication: the described action softens, the negative bans become wishes, and identity — the thing you most need held — is the first casualty, because the model feels less bound to your description of who this person is.
Guidance is a dial with two failure modes, and the working range is the middle band between them. Knowing your tool’s middle by behaviour, not by the number printed on the slider, is the skill.
When High Guidance Vandalises Your Negatives
The vandalism happens through a side door. High guidance makes the model pursue the positive prompt so aggressively that it’ll reintroduce banned things as by-products. You banned text; the prompt says “café menu board”; the model’s obedience to the noun overwhelms its obedience to the ban, and the board arrives with letters — bad ones, which is worse than the clean board a calmer setting might have produced.
You banned other people; the prompt says “busy station”; the crowd arrives anyway, because at high guidance the described world outranks the excluded world. Creators read this as “the negatives did nothing,” raise guidance further in frustration, and tighten the spiral. I know. I’ve been there.
The second vandalism is aesthetic. Strong guidance plus strong negatives tends to produce that airless, cut-out look: subjects separated from environments, edges too certain, light too committed. The model, forbidden from so many of its natural transitions, renders a frame that feels pasted together. If your footage has started looking like assembled clip-art while your settings crept upward over a project, you’re likely watching this — not a decline in the tool.
The fix is counterintuitive and consistent in my experience: when negatives seem to stop working, don’t add more negatives and don’t push guidance up. Lower guidance slightly into the range where text reads as direction rather than law, shorten the negative list to the shot’s real risks, and make the positive prompt carry more of the exclusion work positively — blank sign, empty street, plain surface.
Nine times out of ten the bans start registering again, because they’re no longer competing with an over-amplified description of the very things they were banning. That pairing — modest guidance, short negatives, descriptive positives — is the quiet engine inside my prompting system. And it’s outlived every tool interface I’ve worked in.
Low Guidance And The Washed-Out Look
The opposite failure gets less discussion, probably because it feels gentle. At low guidance, footage often looks pleasant in a vague way: soft, drifting, slightly underwater. Actions half-happen. A character meant to reach for a cup considers it.
The negatives nominally remain set, but with the model loosely bound to text, “avoid” becomes weather it may or may not walk into. Faces drift sooner. Continuity between the frame you approved and the clip you got quietly loosens — and in image-to-video work, where the approved frame is the contract, that looseness is expensive. The strength/freedom side of the panel compounds it: paired with generous departure from the starting image, low guidance hands the model both weak instructions and permission to ignore the frame. The output can still be pretty. It just isn’t your shot anymore.
I see this most often when creators copy settings from a dreamy nature-scenery tutorial into character work, or when a tool defaults to a cinematic-loose mode that flatters demos and betrays production. The tell is QC: the clip is fine, the clip is fine, the clip is fine — and the sequence doesn’t cohere, because each generation drifted toward the model’s taste instead of yours.
Temporal consistency dies by a thousand polite drifts. The remedy is boring: raise guidance into your tool’s obedient middle for any shot carrying identity, action, or story information, and reserve the loose end for atmosphere shots where drift is the aesthetic. Match the dial to the shot’s job, the same way negatives are matched to shot type. There are no project-wide settings in professional work. There are shot-wide settings, chosen on purpose.
A Tuning Order That Respects All Three
Order is the practical heart of this article, because most dial misery comes from tuning in a sequence that hides causes. Here’s the order I run, every time — and the order itself matters more than the starting values.
First, the frame. Approve the starting image before any dial is touched, because no setting repairs a wrong cast, a wrong room, or wrong light — and a strong frame dramatically reduces how hard the dials must work. Second, the words: positive prompt in plain language with exclusions folded in positively where possible, then the short aimed negative set. Third, guidance: set it in the obedient middle of your tool’s range and generate one honest take.
Read that take for obedience specifically — did the action happen, is the subject right, do the negatives register at all? If obedience is weak, raise guidance a step and go again. If the frame feels strained or overcooked, lower it a step. One step, one generation, one read. Fourth, and only then, the strength or motion side: adjust how much energy and departure the shot carries, in small moves, watching that the face and the frame contract survive the added life.
What this order prevents is the classic tangle: changing three things per generation and learning nothing. One dial per iteration is slower per click and dramatically faster per project, because every take answers a question. I also log the final settings beside the approved take in the project notes — where a tool exposes seeds, I log those too, following the same record-keeping instinct as my seed methodology. Settings discovered at midnight and not written down are settings you’ll rediscover at full price on the next project.
Defaults Worth Starting From
Since exact scales differ between tools, I’ll give you posture instead of numbers. And honestly, I’d distrust anyone who gives you universal numbers without asking which tool and version you run. Start guidance in the middle of whatever range you’re given — not the default necessarily, the middle — because manufacturers place the obedient band around there more often than not.
Start the motion or strength side low-to-moderate for anything with a face, moderate for objects and environments, and only climb when a take proves the shot needs more energy than it’s getting. Start negatives short: your base universals plus the one module the shot type calls for. Those three postures together produce a first take that is diagnosable, which is the actual goal.
A perfect first take is luck. A diagnosable first take is a workflow.
From there, let the shot type referee. Close character work: guidance mid-to-slightly-high, motion restrained, negatives structural. Product work: guidance mid, motion minimal, negatives on shape and text. Wide atmosphere: guidance can breathe lower, motion can climb, negatives shift to geometry and flicker.
Write your own shot-type defaults into your prompt template once a project teaches you something, the same way the negative library grows — from QC evidence, not from a settings screenshot in somebody’s tutorial. And revisit them when tools update, because dial behaviour does drift between versions, quietly retuning your defaults underneath you.
The thought to keep from all of this: guidance, strength, and negatives are one system wearing three labels. Your negatives aren’t being ignored out of malice; they’re being outvoted, or under-amplified, or obeyed so hard the footage stiffens. Tune the frame first, the words second, guidance third, and the energy last — in small steps, with honest reads between them. Do that and the dials stop fighting. Because you finally gave them an order to speak in.