September 5, 2026
The 10 mistakes that burn your credits in generative AI
Exploring on a 19 credit model, generating video before validating in stills, regenerating instead of editing, buying prompts: the ten habits that cost the most, and what to do instead.
Frank HoubreFounder of Imaginode
The same ten mistakes, in almost everyone
Most waste in generative AI comes from ten precise habits, all fixable without learning anything technical.
I have been training teams on these tools for a while, and what strikes me most is the repetition. It is never ten different mistakes per person, it is the same ten mistakes in everyone, whatever their starting level.
None of them is a matter of talent or technical understanding. They are reflexes inherited from other tools, or simply things nobody ever said out loud. Once corrected, the same person's monthly budget is divided by two or three, for better results.
They are ranked by what they actually cost, from the most expensive to the most insidious. If you fix only one, take the second, it is by far the one that hurts most.
BalanceMistake 1: exploring on an expensive model
Searching for an idea on a 19 credit model costs nineteen times the price of a 1 credit attempt, for a decision that does not depend on render quality.
The reflex makes sense: I want the best result, so I take the best model. Except that in an exploration phase you are not judging the render. You are looking for a composition, a framing, an idea. A 1 credit image tells you exactly the same thing on those points as a 19 credit one.
Flux Schnell at 1 credit is made for that. Twenty attempts cost 20 credits, about twenty cents. The same twenty attempts on Nano Banana Pro cost 380 credits, almost four euros, for an identical decision.
The rule: the expensive model only comes out once the composition is chosen. Explore at the floor price, finish at the top price. It is the most profitable discipline of them all, and also the easiest to apply.
Mistake 2: generating video before validating in stills
A video shot costs 26 to 96 credits against 1 to 5 for an image: validating the framing as a still first halves or thirds the bill of a film.
Here is the most expensive one, by a very wide margin. You have an idea for a shot, you generate it straight to video. The framing is off. You regenerate. The character's face has changed. You regenerate. Three regenerations at 48 credits, and you still have nothing.
The same three iterations as stills would have cost 9 credits on Flux 2 Klein. Once the framing is right, the image becomes the starting frame of the Video node, and the model only has to animate a composition that is already approved.
That is the whole logic of the storyboard, which I detailed with full figures in storyboarding with AI before paying for video. Remember the ratio: a correction as a still costs thirty times less than a correction in video.
Mistake 3: describing a face instead of showing it
No prompt, not even a four hundred word one, stabilises a face: reference images do it, text never will.
Your character's face changes with every generation, so you add details. Nose shape, eye spacing, the lock of hair falling on the left. You reach four hundred words and the face keeps changing.
Language is too poor for a face. Describe your best friend in writing and hand the text to an illustrator who has never met them: the portrait will be plausible, never a likeness. And every word added about the face is paid for in attention lost on the scene and the light.
What works: a Reference node, three photos, an @mention in the prompt. The model receives images and copies real proportions. The full subject is in keeping a character consistent.
Mistake 4: regenerating instead of editing
When an image is 95 percent right, regenerating it starts everything from zero: an editing model changes the detail and keeps the rest for 2 to 10 credits.
The image is perfect, the jacket just needs to be blue. You change the word in the prompt and hit generate again. You get an entirely different image: different face, different set, different light. And you have lost the good one.
Every generation starts from different random noise. Regenerating is never modifying, it is starting over. Flux Kontext at 6 credits, GPT Image Mini at 2 credits or Nano Banana 2 at 10 take your image as input and change only what you describe.
And remember the second half of the instruction, the one everybody forgets: the list of what must not move. Face, pose, framing, background, light. Without it, the model retouches everything it passes. The full method is in making the same AI image again.
Mistake 5: insisting with a model that cannot do it
Three failures in a row on the same defect mean the model cannot do it: a fourth attempt will change nothing.
The textbook case is text inside the image. You want a brand name on a sign, the model writes almost the right word, you regenerate. Twelve times. Twelve credits for nothing, because that model cannot write.
Never regenerate the same model more than three times on the same defect. Three failures is information: change model. Nano Banana Pro, GPT Image 2 or Recraft V4.1 write short text cleanly and fix it on the first try.
The same reasoning applies to references that are ignored, a port that does not appear, a format that does not exist on that model. It is not a hidden setting, it is a missing capability. The detail by defect is in warped hands and unreadable text.
Mistake 6: generating in one format then cropping
The model composes for the frame it is given: cropping a square into vertical destroys the composition and loses half the pixels, while generating at the right format costs nothing extra.
You generate square because it is the default setting, then the visual goes into a story. You crop, the character ends up cut off, and the background you liked has disappeared.
Format is a dropdown on the node, and it does not cost a single extra credit. In 9:16 the model stacks vertically; in 16:9 it leaves air on the sides. Those are two different compositions, not the same one with extra edges.
And if the same content has to come out in two formats, do not crop: duplicate the node, change the format, keep the same Style and Reference nodes plugged in. The exact sizes for each format are in which format and resolution to generate.
Mistake 7: raising resolution during tests
Ten seconds of Seedance 2.0 cost 120 credits in 720p and 660 in 2160p: exploring at high resolution multiplies the bill by five on shots that will end up in the bin.
In video, resolution is the most brutal price lever in the catalogue, and it is the one people move without thinking because the setting is right there.
The figures: on Seedance 2.0, ten seconds cost 120 credits in 720p and 660 credits in 2160p. On Wan 3.0, thirty seconds are 180 credits in 480p against 720 in 1080p. There is no volume discount.
Yet motion, framing and how a character holds up are perfectly judged in 480p. Validate low, go high only on the version you keep. And with stills the opposite is sometimes true: on Flux 2 Klein and Seedream 5 Lite, the upper tier costs exactly the same, so you may as well take it.
Mistake 8: leaving audio on for shots you will edit together
Each shot then generates its own ambience and the six contradict each other in the edit, for an extra cost of about twenty credits per shot.
Many video models offer an audio toggle, often on by default because a silent video is noticed immediately. On a standalone shot, fine. On six shots meant to be edited together, it is a double loss.
The extra cost first: on Veo 3.1 Lite, eight seconds in 720p cost 29 credits without sound and 48 with. Over six shots that is 114 credits of difference. Then the result: six different ambiences that cut brutally at every join.
Generate silent, then lay a single track over the finished edit. A voice-over, some music, a few sound effects placed where they belong. The film gains the unity that separately scored shots will never have.
Mistake 9: buying prompts
A prompt sold to you was written for another model, another version and another subject: it does not transpose, and it teaches nothing.
Never buy prompts. Those packs of a thousand magic prompts for twenty euros are the most useless product in the sector, for three reasons that fit in three lines.
One: they were written for a specific model, often of an earlier generation, and they give nothing like the same result elsewhere. Two: they describe their author's subject, not yours, and adapting them requires exactly the skill of prompting. Three, and this is the worst: they teach you nothing, so you will buy more next month.
What can be learned is the structure of a prompt, and it fits in one article: writing a prompt that works. And if you want a model to write your prompts, the Text node does it for 1 to 3 credits, with your subject and your art direction.
Mistake 10: keeping nothing
The prompt, the model and the settings that produced a good image are untraceable if the project is not kept: you pay again for work already done.
The most insidious one, because it costs nothing at the time. You find the right combination, you download the image, you close. Three weeks later a variation is needed, and you no longer know the model, the prompt or the settings.
You start over, you get roughly the same thing, never quite the same. Over a month of regular work, that is hours and a whole credit budget spent twice.
On the canvas, every node keeps its generations as thumbnails, and behind each thumbnail the parameters screen gives the engine, the prompt word for word, the settings, the camera, the inputs, the cost and the date. Keep the project rather than the file, it is free and it will save you more than once.
The right question to ask before every click
Ask yourself what this generation has to teach you: if the answer is a composition decision, the 1 credit model is enough.
All these mistakes come down to a single missing habit, and it takes two seconds: before clicking, ask yourself what this generation has to teach you.
If it is a decision about composition or framing, take the cheapest model, at low resolution, as a still rather than a video. If it is the final render, bring out the good model and raise the resolution. Those two situations call for neither the same model nor the same budget, and confusing them is the root of the problem.
The exact price appears on the Generate button before every click, which makes the trade-off immediate. And if you want to price a whole project before starting, the public calculator does the total, or the canvas assistant builds you a complete flow fitted to your budget.