Leonardo AI makes an appealing promise: generate a product photograph, adapt it for a website, create an advertisement, edit a mistake, and animate the result without moving between several applications.
I wanted to know how much of that workflow I could trust with a recognizable product rather than an attractive image that only looked right at a glance.
The answer turned out to depend heavily on the model, the kind of image, and how strictly I needed it to preserve a brand.
Quick Verdict
For concept work, mascots, mood-setting imagery, and quick visual exploration, I would comfortably use Leonardo AI again.
For accurate ecommerce products, logos, or tightly controlled campaigns, I would use it as part of a workflow that includes inspection and conventional editing. I would not upload a real product photograph and assume an attractive generated result faithfully represents the item a customer will receive.
The platform’s strongest practical advantage is choice. I could stay in one interface while switching between Lucid Origin, GPT Image 2.5 Flare, Nano Banana, MiniMax H3, and Seedance 2.5.
That flexibility matters because the performance differences weren’t subtle. An 8-token composition from Lucid Origin took 10 seconds but reversed the duck’s direction, whereas a 70-token Flare result took 22 seconds and obeyed the detailed composition.
My biggest reservation is the hidden cost of inspection and correction. A single generated product image may be cheap, but several failed iterations, a 40-token edit, and a premium 1,080-token video can turn a small creative task into a much larger spend.
A paid tier and a sizable balance made experimentation comfortable; the usable asset still needed judgment.
| Test | Score (Out of 10) | Why This Score |
|---|---|---|
| Product photography | 5.5 | Attractive landscapes, but both Lucid attempts redesigned reference details. |
| Website hero image | 7.0 | Useful wide layout and negative space; product accuracy remained compromised. |
| Social advertisement | 8.0 | Flare delivered exact text; the inexpensive Lucid version required significant correction. |
| People and realism | 4.0 | Two attempts obscured the hiker; facial realism could not be evaluated. |
| Illustration and character | 9.0 | Fast, appealing fox mascot, with minor prop and clothing inaccuracies. |
| Precise prompt following | 8.5 | Flare met the specified constraints; Lucid reversed the duck’s direction. |
| Editing and consistency | 7.0 | Strong background edit; imperfect text repair and product preservation. |
| Image-to-video | 7.0 | Attractive motion, but brand lettering and compass continuity failed. |
| Overall | 7.2 | Flexible image-first workflow; costly corrections and strict-branding reliability concerns. |
Leonardo AI Pros and Cons
- Fast Lucid Origin image generation
- Strong illustrated mascot quality
- Flexible model choices in one workspace
- Accurate text with GPT Image Flare
- Useful natural-language image editing
- Effective two-times image upscaling
- Accessible image-to-video creative workflows
- Reference products changed shape and branding
- Premium video renders were costly and slow
How I Tested Leonardo AI
I tested on October 9, 2026, with an Essential account showing 8,500 Fast Tokens at the start, a 25,500-token rollover-bank capacity, and private generation enabled.
The plan screen listed a $12 monthly Essential price excluding tax, and my balance renewed on November 7.
The hands-on work covered six image assignments, two editing operations, one 2× upscale, and five short image-to-video clips.
The model selector was part of the test: each tool’s controls, reference compatibility, dimensions, and token price affected what I could reasonably produce.
The reference was a square studio picture of a dark navy-blue HARBORTRAIL hiking backpack with a white mountain-marked black patch, a front criss-cross bungee arrangement, buckles, side mesh pockets, a lower zippered compartment, and a separate black flashlight on the right.
The backpack was a fictional controlled product, not a physical item whose exact dimensions I verified with a camera. Keeping the same source allowed me to notice when a generated image invented another logo, moved the pocket, introduced zippers, shifted the material color, or re-created the flashlight against instructions.
The six assignments were practical rather than arbitrary benchmarks:
- Put the backpack at a mountain campsite
- Make a wide hero banner with negative space for copy
- Produce a square promotional ad with exact text
- Show a human interacting with the backpack
- Create an original fox mascot for the brand
- And construct a studio scene with a red mug, yellow rubber duck, and blue vase in a required order
They tested different purchasing needs, from ecommerce photography to illustration and precise prompt following.
My assessment categories were ‘usable unchanged’ when the result served its intended purpose without editing, ‘minor correction needed’ for an otherwise useful asset with a small fix, ‘substantial rework needed’ when composition or branded details required meaningful repair, and ‘failed’ when the main requested subject or action wasn’t delivered.
I distinguished visual appeal from factual product accuracy. A beautiful but redesigned backpack could be a good concept and still be a bad product advertisement.
Generation times were wall-clock values recorded for each attempt. They do not include my prompt writing, download, inspection, or manual correction time unless specifically stated. Token figures are actual observed balance changes or prices shown by the interface, with the source of each clearly identified.
One accidental four-image Flare batch was charged 282 tokens; I retained it in the spending record rather than treating it as a single-image benchmark. None of the sample outcomes is a general success rate across every Leonardo model or prompt.
The model settings were not identical across every comparison.
For example, Lucid Origin used Fast and Dynamic at 1024 × 1024 in most still-image tests, whereas GPT Image 2.5 Flare used High quality, and the social-ad comparison used different image-reference modes. The first person-interaction attempt also inadvertently had Prompt Enhance on Auto rather than Off, while the retry was Off.
In the three-object comparison Flare’s screenshot likewise showed Auto. Those differences limit strong head-to-head claims.
1. Product Photography Test
I began with product photography because a store cannot afford to invent features on the item it is selling. The reference was the original navy HarborTrail backpack, and my buying question was simple: could Leonardo place it outdoors without altering the product?
I selected Lucid Origin, Dynamic style, Fast mode, a 1024 × 1024 canvas, one generation and Content Reference. The first image took 21 seconds and cost 10 tokens.
The initial picture looked convincing from a distance. The backpack stood upright on a rock in front of distant mountains, with plausible fabric and shadows.
But close inspection revealed a redesign. The lower bag area acquired extra HARBORTRAIL lettering, the front compartment and bungee arrangement changed, and straps and proportions drifted.
The prompt requested a background transformation, not a new backpack. That distinction made it unsuitable as a product-accurate image even though it looked polished.
I repeated the assignment with a stronger preservation prompt, specifying the original patch, zipped pocket, mesh sides, buckles, straps and bungee arrangement, and explicitly prohibiting new branding or an added flashlight. Another 10-token output followed.
Instead of restoring the exact original, the new image again altered the backpack’s pocket structure and introduced visual details, including additional red and small label elements. Repeating ‘exact same backpack’ did not give me an exact copy.
The two outputs demonstrate why a reference control should not be mistaken for a product-locking tool. They preserved the broad category, silhouette, and color family while changing details that would matter to a customer.
I could use either to pitch an outdoor campaign concept, but I would not use either as the main product listing photograph. Their combined generation charge was 20 tokens; neither was an unchanged, product-faithful deliverable.
2. Website Hero Image Test
A useful product photograph is only the starting point for a website. A hero banner must also leave space for a headline, so I took the same backpack into a wide-layout assignment to see if Leonardo could solve composition without another round of product redesign.
I kept Lucid Origin and Content Reference, switched to 16:9 at 1376 × 768, and requested a golden-hour mountain setting with quiet negative space for copy.
Leonardo placed the backpack against an imposing rock formation with a blurred alpine backdrop and a relatively quiet patch on the left. The landscape and warm light made a believable premium outdoor mood.
Importantly, this image did not contain the marketing headline itself; the planned text was to be added separately. That was a reasonable division of labor for a website designer who needs to retain full control over copy and typography.
The problem was still product fidelity. The bag’s orientation, lower area, front cord arrangement and added equipment differed from the original reference. The picture included a dark cylindrical item hanging by the side that resembled the flashlight explicitly excluded from the request.
The background was serviceable, but the campaign asset no longer showed the reference product exactly. The recorded charge was 10 tokens.
For a site using a fictional brand or conceptual prototype, I could place conventional text on top of this image and move forward. For a store selling a specific hiking pack, I would need either manual compositing with the verified product photograph or a much more controlled editing process. Wide layout success does not cancel out object-identity failure.
3. Social Advertisement and Text Test
The hero image left enough room for copy, but there was no guarantee Leonardo could write that copy accurately inside an ad. That mattered because a wrong headline or discount can make an otherwise appealing campaign image impossible to publish.
I requested a square HARBORTRAIL advertisement with the exact headline ‘READY FOR THE TRAIL’ and the exact offer ‘15% OFF,’ without extra promotional words or altered branding.
Lucid Origin generated a mountain product ad in 19 seconds for 10 tokens. The main discount, 15% OFF, was present, but READY FOR THE TRAIL was absent. Random-looking vertical lettering appeared along the left edge; the backpack itself had misshapen branding and changed physical details.
A circular HARBORTRAIL badge added to the top corner looked promotional, but the smaller markings did not survive close inspection. That combination was not suitable for a paid campaign.
I then switched to GPT Image 2.5 Flare, which used a different reference mechanism and displayed High quality.
Leonardo warned that the earlier Content Reference was incompatible, so I changed it to the supported Image Reference before proceeding. The intended single-image output was charged 73 tokens and took 25 seconds.
There was also an accidental four-image generation charged 282 tokens immediately beforehand, later explained by my inadvertently leaving the image count at four.
I separated those actual charges in the ledger rather than attributing all 355 tokens to the final successful image.
Flare’s chosen ad clearly rendered READY FOR THE TRAIL and 15% OFF, preserved a recognizable mountain-patch HARBORTRAIL backpack, and gave the discount a clean contrasting shape.
The composition was much closer to what a marketer would expect from a completed graphic.
The backpack was not geometrically identical to the source, but it kept substantially more of the intended visual identity than the Lucid Origin advertisement. I classified the Flare output as usable unchanged for a fictional campaign mockup, not proof of exact product fidelity.
There is a useful nuance here: Flare did not produce the superior image only because I wrote the same prompt and swapped the model. Its reference mode, quality controls, and prompt formulation were different.
The result supports a narrower conclusion that within this actual workflow, the higher-cost Flare generation followed the requested campaign text and layout better. It does not establish that Flare will win every text or advertising test.
4. People and Realism Test
After testing the backpack alone in product shots and advertising layouts, I wanted to add a person. Human-product interaction raises a different purchasing question: can a business get a believable lifestyle shot, with the subject visible and the hands actually contacting the product, without organising a photoshoot?
I asked Lucid Origin for one adult hiker interacting naturally with the same backpack against mountains. The first Fast 1024 × 1024 result cost 10 tokens and took 12 seconds; I needed a visible person as well as a recognizable bag.
Instead of showing an identifiable person interacting with the product, Leonardo filled the frame almost entirely with the backpack. A small hand appeared on either side, as though someone stood behind it, but no face or body was available to evaluate.
The bag’s front layout changed again. I could not honestly call that an anatomy failure based on facial distortion, because a face wasn’t generated; it was a failure to obey the composition and subject-visibility request.
For the one permitted retry, I made the person-placement instruction explicit: show the adult from roughly mid-thigh upward beside the backpack, with one hand holding its top handle and the other steadying the side.
I also corrected Prompt Enhance from Auto to Off. The second 10-token result took 12 seconds and again concealed nearly the entire person behind the bag.
Forearms and hands were visible at the top, but the face and most of the body disappeared. The generated backpack had further differences in pocket treatment, front cords and logos.
These two outcomes do not justify a sweeping claim that Leonardo cannot generate realistic people.
They show a narrower but commercially important weakness in the Lucid Origin reference-driven workflow: my instructions to show both the product and an unobscured person were not followed. After two attempts and 20 tokens, I had no usable person-product photograph and stopped, as planned.
5. Illustration and Character Test
Two failed person-product compositions made me curious about a different strength: creating something from scratch rather than preserving an existing object. For a small brand, an original mascot might save more concept-development time than another imperfect product photograph.
I asked Lucid Origin for a 3D-style HarborTrail fox with orange fur, a cream muzzle and chest, emerald-green eyes, dark ear tips, one white-tipped tail, a navy vest with two pockets, a mustard-yellow neckerchief, and a small silver compass.
The fox was to stand on a mountain trail. I did not upload a reference image.
This was Lucid Origin’s strongest image result in the series. For 8 tokens and an 8-second generation, it produced a friendly character with consistent proportions, a clean 3D animated-film appearance, soft daylight, and a legible expression.
The orange fur, green eyes, cream face, dark vest and yellow scarf were immediately recognizable. The fox had two arms, two legs, one tail and two ears; I did not see extra anatomy or disruptive visual noise.
The compass was less exact. It resembled a small round medallion, with few readable details to verify it as a navigation instrument.
The image also added shorts, an accessory that had not been requested and was actually excluded in the original instruction’s ‘no other clothing’ clause. That did not destroy its usefulness as an original fictional mascot, but it is still a prompt-adherence miss rather than something to overlook because the art is appealing.
The real value here is speed to a plausible character concept. A designer could take this illustration into mood boards, prototype packaging, or social posts immediately.
Establishing true character consistency across many distinct scenes would require repeated controlled generations that I did not run; a single polished pose is not evidence that every future expression or angle will match. I later reused this same image for upscaling and video, allowing a limited view of identity preservation.
6. Precise Prompt-Following Test
The fox showed how convincing Leonardo could be when it had freedom to create. My final still-image assignment reversed that freedom.
I gave the model a small scene with rules that could be checked exactly, so I could separate attractive styling from instruction-following.
I requested a photorealistic square studio image with exactly three objects in one horizontal row: a red ceramic mug on the left with its handle facing left, a yellow rubber duck in the centre facing right, and an empty cobalt-blue glass vase on the right. No extra props, labels or text were allowed.
GPT Image 2.5 Flare generated the studio arrangement in 22 seconds for 70 tokens. All three objects appeared once, in the requested left-to-right order, with the correct colors.
The mug handle faced left, the duck faced right, and the vase was empty.
The objects were separated cleanly and showed convincing material qualities: glossy ceramic, rubber-like yellow plastic and transparent colored glass. It passed the 10 original checklist items and the additional directional requirements.
I used the exact same prompt with Lucid Origin, Prompt Enhance Off, Fast mode, Dynamic style, and 1024 × 1024 dimensions. That 8-token attempt took 10 seconds. It correctly created three objects, kept the stated colors and broad positions, and gave the mug a left-facing handle.
However, the duck faced left rather than right. The vase also appeared more opaque and the composition left a large blank area above the objects.
The initial broad ten-item checklist missed duck orientation, so I recorded that specific failure separately instead of retroactively changing the benchmark score.
This side-by-side was one of the cleanest tests of the tradeoff between cost and compliance. Lucid Origin was almost nine times cheaper in tokens and substantially faster for the single image, but it missed an explicit directional instruction.
Flare gave me an image I could use immediately for the invented studio assignment. Because Flare’s interface showed Prompt Enhance Auto while Lucid was Off and their quality controls differed, the experiment remains a practical comparison, not a perfectly isolated model trial.
Product and Brand Consistency
The backpack sequence exposed the difference between visual continuity and identity preservation.
Across the square mountain photograph, correction attempt, 16:9 website hero, and square social image, most outputs remained recognizably ‘a dark hiking backpack in an outdoor setting.’ That is not the same as being the exact pack in the studio reference.
The number of cords, stitching geometry, strap placement, patch location, lower pockets and extra lettering changed repeatedly.
The most revealing defect was additional brand text. The first mountain image showed another HARBORTRAIL wordmark on the lower front, a detail absent from the initial studio reference.
A later attempt added another conspicuous label; the square Lucid advertisement mixed a correctly readable 15% OFF offer with garbled incidental lettering. In a real product campaign, this could create misleading product features or trademark problems even when the model gets the dominant brand word right.
Flare’s square campaign graphic was closer to the source packshot, but its success was also relative.
It should not be used as proof of exact stitching, dimensions or material matching. The edited desert image later preserved much of its immediate source’s geometry, yet that source was itself a generated interpretation of the original backpack.
An edit can be faithful to the wrong predecessor. I therefore compared each stage to both its immediate source and the initial reference.
The fox illustrated a different type of consistency. The single-source original contained distinctive orange fur, green eyes, one white-tipped tail and a navy vest.
Upscaling kept these major traits, as did both short animations, although their hand-held compass did not stay consistent. For character-led media, a creative model can preserve the general mascot identity while still changing the state of a small prop from one frame to the next.
My practical approach would be to use these tools for ideation and background variation, while keeping a verified product image and a written design checklist outside the generator.
If the exact item must be displayed, use a non-generative layer of the real photographed product when possible. For brand consistency, don’t judge the image only at thumbnail size; zoom in on labels, straps, seams, and recurring character accessories.
Editing and Upscaling
After the six assignments I tested two natural-language edits and one 2× upscale. This was important because an image-generation platform is more useful if I can correct one defect without paying for and reviewing an entirely new composition.
I used the existing Library images and stopped at one edit attempt per case.
For the text correction I opened the failed Lucid Origin social advertisement, selected the prechosen Nano Banana editor, and instructed it to add READY FOR THE TRAIL in bold uppercase letters while preserving the backpack, background and 15% OFF. The operation cost 40 tokens. The requested headline was inserted with the exact words, while the original offer and much of the backpack stayed in place.
However, the new lettering overlapped the upper backpack and circular brand badge. Existing gibberish along the left and defective smaller logos also remained. Fixing one omission did not clean the entire graphic.
That distinction matters when you estimate a creative task’s true cost. The operation succeeded at adding the requested phrase, but the result still needed substantial manual layout and lettering repairs before publication. I did not run a second correction, so I cannot estimate how many further edits would have been needed to make it production-ready.
For background replacement I selected the first generated mountain backpack, asked Nano Banana to replace only the scenery with warm desert dunes, and explicitly required the bag’s proportions, logos, pockets, and cords to remain unchanged.
The edit took 15 seconds and charged 40 tokens, confirmed by the balance dropping from 7,959 to 7,919. The desert was convincing.
The backpack stayed upright, in the same position, with its major features largely intact. The original foreground rock remained and the bag’s existing invented label details persisted.
That was a stronger editing result. It did the main job without visibly rebuilding the entire pack.
Still, the successful preservation was preservation of the flawed mountain output, not of the true studio product. I would classify it as usable for a fictional background-change demonstration, not as a verified merchandise photograph.
I also opened the fox illustration in Leonardo’s Upscale control, selected Pro, Precise, 2×, and left Fix AI Image Artifacts Off.
The cost was 18 tokens; processing took less than ten seconds. The output dimensions shown in the Library increased from 1024 × 1024 to 2048 × 2048, or approximately 4.1 megapixels.
The fox’s clothing, expression, green eyes and tail remained recognizable, with a cleaner enlarged presentation. A larger pixel count is not evidence of newly recovered photographic truth; the benefit here was practical output size for the illustration.
In all three operations, I needed to understand Leonardo’s controls. The image preview gave me prompt editing, Remove Background, Remix, Upscale and Create Video choices, and the preselected editing model affected price.
I did not test Magic Layers, Canva’s separate editing workflow, or background-removal quality. These are visible options, not measured successes.
Image-to-Video Testing
I made five six-second image-to-video clips from the images already generated, without creating additional source art. Three used MiniMax H3, and two used Seedance 2.5.
The goal was deliberately narrow: test smooth motion, stability of the input image, text persistence, character identity and the handling of a small compass.
I chose one backpack product scene, one fox mascot and one social advertisement for MiniMax, then compared Seedance on the advertisement and fox.
MiniMax H3 generated the desert backpack animation in 12 seconds with a displayed 90-token cost. The model delivered a smooth push-in toward the stationary product; the desert remained recognizable, and the backpack’s larger features stayed consistent.
The requested blowing sand was subtle to the point of being difficult to isolate. This was the most immediately useful product clip of the three MiniMax generations.
The MiniMax fox clip took 13 seconds and cost 90 tokens according to the surrounding balance history. Its blinking, tail motion and facial styling were appealing, but the compass briefly disappeared and reappeared in the other paw.
That movement broke object continuity even though the character remained recognizably the same fox. For a brand mascot holding a prop, that is exactly the kind of small error a campaign reviewer should catch.
The MiniMax social advertisement took 14 seconds, again at 90 tokens. It kept the large headline READY FOR THE TRAIL and the 15% OFF promotional message readable while adding a restrained camera push-in.
The backpack’s smaller HARBORTRAIL patch became less reliable as the image animated. The result was useful as a fictional campaign concept, though I would correct the fine branding before a real paid placement.
Seedance 2.5 changed the economics significantly. I initially saw a 1,752-token price for a six-second 960 × 960 HD generation. Lowering resolution to Standard 640 × 640 reduced the displayed price to 1,080; switching Audio Off did not lower it further.
I used that lower-cost configuration for the two actual Seedance clips. The export was therefore smaller than MiniMax’s 768 × 768 output, a notable limit on any quality comparison.
Seedance’s social-ad clip took 5 minutes 33 seconds, versus MiniMax’s 14 seconds for the same source and motion request. Both kept the large campaign words legible.
Seedance showed noticeable camera movement and cloud animation, but the increasingly large backpack crowded the headline’s layout toward the end. The result did not justify a categorical claim of higher quality for the higher price.
The Seedance fox took 3 minutes 59 seconds, versus 13 seconds with MiniMax. The animation preserved the fox’s general face and clothing, and its tail moved more noticeably. But the compass again changed hands and its appearance shifted; the tail also behaved unnaturally at moments.
Neither model passed the object-continuity requirement for the fox. I stopped at five clips rather than spending another 1,080 tokens on a sixth backpack comparison that was unlikely to answer a new purchasing question.
Generation latency matters here as much as pixel count. For repeated short social concepts, 12–14 seconds per MiniMax clip can make rapid revision practical.
Four to five and a half minutes per Seedance clip changes that rhythm, especially when a defect may require a full re-render. These timings are observations from single runs, not guarantees of service-wide speed.
Which Leonardo AI Models Worked Best?
Leonardo AI offers dozens of image and video models, including its own tools and third-party generators.
That choice is useful, but the best model depends on what you need to publish: a quick concept, an accurate advertisement, an edit, or a video.
I tested five models in my HarborTrail assignments, not the entire catalog. The comparison below shows what each delivered, how much time and how many Fast Tokens it used, and when I would pay for it again.
| Model | Best Observed Use | Time & Fast Tokens | Buying Implication |
|---|---|---|---|
| Lucid Origin | Quick concepts and the fox mascot | Fox: 8 sec / 8 tokens; three-object scene: 10 sec / 8 tokens | Fast and inexpensive for ideas. The backpack, hiker visibility and duck direction were unreliable for exact briefs. |
| GPT Image 2.5 Flare | Social ad text and precise object layout | Ad: 25 sec / 73 tokens; three-object scene: 22 sec / 70 tokens | Higher token use, but produced usable fictional ad copy and correct placement on the intended attempts. Still check product details. |
| Nano Banana | Background replacement and specific image corrections | 40 tokens per edit; background edit: 15 sec; headline edit time not recorded | Useful for scene changes. The headline edit overlapped the design and left earlier lettering defects, so manual checking was needed. |
| MiniMax H3 | Fast six-second video concepts | 12–14 sec / 90 tokens per clip | Better speed and token value in my video tests; fine branding and the fox’s compass still needed review. |
| Seedance 2.5 | More prominent video movement | 3 min 59 sec–5 min 33 sec / 1,080 tokens per clip | Much slower and costlier at the tested settings, with no decisive improvement in compass continuity. |
Token Usage and Cost of a Usable Asset
The Essential account started with 8,500 Fast Tokens and ended with 5,471, a confirmed reduction of 3,029 tokens across the entire session. That includes the accidental four-image Flare batch, unsuccessful generations, editing, upscale operations, and video renders.
It does not represent the price of just one finished advertisement. Keeping the full ledger exposes what a real experiment costs, including mistakes.
The image-generation phase consumed 501 tokens, broken down as follows:
- Lucid Origin commercial images: 40 tokens for two packshots, one hero image, and one original social ad.
- Accidental Flare batch: 282 tokens for four images.
- Intentional Flare advertisement: 73 tokens for one image.
- Lucid person attempts: 20 tokens for two generations.
- Fox mascot: 8 tokens.
- Flare three-object composition: 70 tokens.
- Lucid three-object counterpart: 8 tokens.
The editing and upscaling stage consumed another 98 tokens: 40 for the Nano Banana headline edit, 40 for the background replacement, and 18 for the Precise Pro upscale.
The first editor’s time was not recorded, the background edit took 15 seconds, and the upscale took less than ten seconds. Those were separate actions, not invisible free corrections to the original images.
Video generation accounted for the largest share of consumption. The three MiniMax H3 clips used 270 tokens in total at 90 each, while the two Seedance 2.5 clips consumed 2,160 tokens at 1,080 each.
| Operation | Tokens consumed |
|---|---|
| Image generation | 501 |
| Editing and upscaling | 98 |
| MiniMax H3 videos (3) | 270 |
| Seedance 2.5 videos (2) | 2,160 |
| Total | 3,029 |
The ledger reconciles exactly with the account balance, without assuming any missing charges.
For a usable promotional asset, the comparison is more revealing than the list price per generation. The Lucid square ad at 10 tokens needed work; its 40-token Nano Banana headline correction still left a defective layout and branding. That path spent 50 tokens without reaching a production-ready ad.
Flare’s intended single-image advertisement cost 73 tokens and yielded a usable fictional campaign mockup. The accidental 282-token batch was a separate real expense, not evidence that one Flare generation always costs 355 tokens.
On Essential, the session’s 3,029 tokens amounted to approximately 35.6% of the 8,500 monthly allocation.
I would not convert that directly into a fixed cash price per usable image because a subscription buys other features, and token rates change with model, quality, resolution, and duration. Token cost per attempt is measurable here; true business cost also includes an operator checking the image and correcting errors.
Leonardo AI Pricing and Plans
Leonardo AI’s official pricing page, which I checked for this review, lists four individual plans: Free, Essential, Premium, and Ultimate.
Prices start at $12/month for Essential and reach $60/month for Ultimate, excluding tax. These are monthly billing prices, with up to 20% off advertised for annual subscriptions.
Here’s how the plans compare:
| Plan | Monthly Price | Fast Tokens | Rollover Bank |
|---|---|---|---|
| Free | $0 | 150/day | — |
| Essential | $12 | 8,500/month | 25,500 |
| Premium | $30 | 25,000/month | 75,000 |
| Ultimate | $60 | 60,000/month | 180,000 |
I tested the Essential plan, so I can speak to what its $12 monthly subscription actually provided. I could enable Private Mode, generate images with premium third-party models, create image-to-video clips, and edit or upscale previous outputs. Leonardo also displayed model-dependent token costs before generation.
Essential supports two simultaneous generations, unlimited personal collections, and a higher image-reference allowance than Free. I completed my testing with 5,471 of the original 8,500 Fast Tokens remaining, despite running multiple image and video experiments.
Premium increases concurrency to three simultaneous generations and adds unlimited relaxed image generation for selected models. Ultimate offers more concurrent generations and selected relaxed video generation alongside its larger token allowance.
However, unlimited does not mean every model and generation is free of token charges. It applies to specified models and relaxed queues. For example, I encountered a MiniMax model card displaying an Unlimited label, yet its Generate button still showed a 90-token charge under my configuration.
Leonardo also advertises team and API options, but I did not purchase or test those services.
For someone testing Leonardo for the first time, I’d recommend starting with Essential. Its 8,500 monthly Fast Tokens were sufficient for my experiments, and the lower subscription cost makes it easier to establish how many usable assets you can produce before committing to a higher tier.
Keep in mind that annual promotions, regional taxes, and model-specific token charges can change. Check the current checkout total and generation costs before subscribing.
Biggest Limitations I Found After Using Leonardo AI
- Product fidelity was inconsistent: repeated Lucid Origin mountain scenes changed the reference backpack’s pockets, straps and lower branding, limiting their usefulness for real merchandise pages.
- Prompt precision varied by model: Lucid missed the requested ad headline and later reversed a duck’s facing direction, even though the overall pictures looked convincing.
- Editing repaired the requested change, not the whole asset: Nano Banana inserted a headline but overlapped the pack and left older gibberish lettering untouched.
- Human-product interaction remained unresolved: two Lucid attempts concealed almost the entire hiker, so neither provided a useful portrait or product-lifestyle composition.
- Animation continuity was fragile: both video models changed the compass between the fox’s paws despite instructions to keep its handling consistent.
- Premium video latency and cost were substantial: Seedance used 1,080 tokens per six-second clip and took several minutes, compared with 90-token MiniMax renders in seconds.
Who I Would Recommend Leonardo AI To
Based on my testing, I would recommend Leonardo AI to creators who need visually varied concepts quickly and are comfortable reviewing and correcting the results. It is particularly useful for:
- Freelancers and graphic designers: Creating original mascots, exploring visual styles, and generating backgrounds or initial design concepts.
- Marketers and social media managers: Producing promotional mockups, campaign visuals, and short AI-generated videos for social media.
- Small creative teams: Experimenting with different image and video models in one workspace without switching between separate services.
- Independent content creators: Generating illustrations, scenic images, and supporting visuals for personal or commercial content.
However, I would hesitate to recommend Leonardo as the sole production tool for ecommerce catalogs requiring exact product details, consistent branding, or characters that must handle objects accurately across multiple scenes. My tests showed that these requirements could still produce unreliable results, even after additional generations or edits.
For those projects, I would combine Leonardo with verified product photography, conventional image editing, or careful compositing. Canva or another image editor may also be necessary for precise typography, positioning, and final export layouts.
Leonardo is most convincing when I treat it as a creative production assistant rather than a replacement for every stage of professional design.
Conclusion
Leonardo AI succeeded most when I let it create, not when I asked it to reproduce something precisely. The original fox mascot needed only eight seconds and eight tokens to produce a convincing commercial-style character. The three-object Flare image followed an exact brief.
Nano Banana gave me useful edits. MiniMax H3 made short animations with very little waiting. Taken together, those results make Leonardo a versatile place to develop brand visuals.
The limitation became clear whenever the exact identity of the product or a handheld object mattered. Beautiful backpack scenes repeatedly changed the pack. A headline fix did not clean all the lettering.
Two people-and-product prompts failed to show the person. Both animated fox clips moved the compass to the other paw. These aren’t reasons to dismiss the platform; they’re reasons to treat its outputs as reviewable creative work rather than automatically finished assets.
My Essential account spent 3,029 of its 8,500 monthly Fast Tokens across the full project and finished with 5,471. Most spending came from the two Seedance videos, not the dozens of seconds spent generating ordinary images.
If you need practical advertising assets, calculate the tokens and human attention required to reach an acceptable result. The fastest inexpensive image is not always the cheapest finished asset, and the most expensive model did not always fix the important problem.
