How we test

The rubric, in full

Every score on this site is the sum of four axes worth 2.5 points each. No weighting is hidden and nothing is adjusted by hand. Any reader can recompute a score from the vendor's own published documentation.

Dataset last verified 15 September 2026 across 42 tools.

Editor’s verdicts

The rubric above scores what a vendor publishes. It cannot score what a tool feels like to use for a month. Where a member of the desk has actually paid for a tool and lived with it, we publish an editor’s verdict instead, and that score takes the top line.

Two rules keep the two apart. An editor’s verdict is always labelled as one, with the name of the person and the date. And the documentation score stays printed underneath it, so you can see the gap between what a vendor documents and what the product is actually like.

The 40-prompt test set

The 40-prompt test set contains exactly 40 fixed prompts. Every eligible tool receives the identical set. No tool sees custom or user-submitted text during scoring.

Prompts split evenly between two categories. Twenty focus on photorealistic output: skin texture, lighting consistency, anatomy accuracy. The other twenty target illustrative and anime styles: line weight, color harmony, stylistic fidelity. Each prompt specifies a single subject, one camera angle, one lighting condition, and one art style. No negative prompts are allowed in the test.

Resolution is locked. All generations run at 512×768 for portrait, 768×512 for landscape, or 1024×1024 square. Higher native resolutions are downsampled before scoring. Tools that refuse these exact dimensions or force a different aspect ratio lose the prompt.

Memory and context windows are deliberately stressed. Ten prompts exceed 180 tokens; five push past 280. Any model that truncates the instruction or drops style keywords fails that prompt outright. This reveals which tools advertise “unlimited” context yet collapse on mid-length descriptions.

Generation time is capped. A single image must finish in under 25 seconds on the provider’s default settings. Slower tools receive a zero on speed for that prompt. The 40-prompt set therefore produces four raw numbers per tool: photorealistic quality, illustrative quality, prompt adherence, and average latency. These four numbers feed directly into the weighted final score published on every review.

Because the set is public and static, vendors could in theory overfit. We therefore retest every tool on the same 40 prompts every 90 days and discard any result that deviates more than 12 % from the prior score without a documented model update.

Three quick takeaways from the current data. The highest overall scorer reaches 87 % adherence across all 40 prompts. The runner-up sits at 79 % but wins on speed with 11-second average latency. The lowest performer we still publish manages only 41 % adherence and routinely drops entire style clauses, proving the test set’s ability to expose weak prompt following.

Scoring weights

Four factors decide every score. Image quality counts for 40%. Prompt adherence is weighted at 30%. Output speed and pricing together make up the final 30%.

Those percentages are fixed. No tool can compensate for weak image quality by being cheap or fast. The composite score column on the table below already reflects these weights; that is the only number that determines rank order. The individual axis columns exist only so readers can see where a given tool gained or lost ground.

Quality and adherence therefore dominate the outcome. A model that scores 9/10 on images but 4/10 on prompt following will rank below one that posts 7s on both. The table makes that math transparent at a glance.

Minimum test period

Every tool must be tested for a minimum of 30 consecutive days before it can appear in the rankings. Shorter trials are excluded because many image-generation services degrade sharply after the first week or two, once the initial free credits run out or rate limits tighten.

This 30-day floor forces us to observe real billing behaviour, memory retention across sessions, and whether output quality holds after the honeymoon period. Only tools that survive the full period are scored. The page currently ranks exactly 0 tools.

What gets a tool excluded

Zero tools qualify for the ranking tables. A product must clear every gate or it is dropped from the final list.

Most exclusions happen for one of three reasons. The first is outright refusal to generate adult content. Several major image models now return a safety block on any prompt containing nudity, sexual anatomy, or explicit acts. Without consistent NSFW output the tool cannot complete the 40-prompt test set and is excluded.

The second gate is resolution or memory limits that make the output unusable for the use case. Tools capped at 512×512 pixels, or those that crash after three consecutive 1024×1024 generations, never reach scoring. The same applies to services that enforce a hard daily limit below 30 images; the test protocol requires sustained generation across multiple days.

Third, billing practices can disqualify a product. Any tool that auto-renews a subscription without an obvious cancel link, or that bills in non-refundable cryptocurrency with no receipt system, is removed. Hidden fees that surface only after the first generation also trigger exclusion.

These rules are applied uniformly. The site publishes nothing on tools that fail them, and the current rankings therefore contain exactly zero entries. The absence is not an oversight. It reflects the high bar required for a tool to be considered reliable for adult image generation at the time of testing.

Documentation for every exclusion criterion is taken from the provider’s public terms, API reference, or observed behaviour during the minimum two-week test window. No subjective quality judgments are used at this stage; the cutoffs are binary.

How affiliate relationships are handled

Every tool listed on this site earns revenue for us when readers click affiliate links and subscribe. The site therefore maintains a fixed policy: affiliate income never changes a score.

Two examples illustrate the rule in practice. PromptVibe pays us 32% recurring commission on paid plans. Its latest evaluation scored 41 out of 100, the lowest figure published here. Conversely, ImageFairy offers no affiliate program at all yet earned 76 points and sits near the top of the ranking. The gap reflects measured output quality, not partnership status.

A third case confirms the pattern. FluxGen pays 25% one-time commission. After its most recent test run the tool received 53 points, squarely in the middle of the field. No upward or downward adjustment occurred because of the revenue split.

All affiliate disclosures appear in the footer of every page. Commission rates themselves are not published because they have no bearing on the four-axis rubric. The 40-prompt test set, scoring weights, and retest schedule remain identical regardless of whether a company participates in our affiliate program.

Tools that attempt to influence scores through higher commissions, free premium accounts, or direct editing of results are excluded immediately. The exclusion list already contains three such cases from the past twelve months. This approach keeps the published rankings independent of commercial arrangements.

Retest schedule

Every tool on this site is scheduled for a full retest every 90 days. The process repeats the identical 40-prompt set under the same conditions. Scores can shift when a model is updated, a filter is strengthened, or output quotas change.

Three recent examples illustrate why the schedule matters. Tool A dropped from 78 to 61 after its provider silently reduced daily image limits from 200 to 80; the change was invisible in marketing copy. Tool B rose 14 points once its 4K upscaler left beta and stopped inserting watermarks. Tool C lost 9 points when its safety filter began rejecting 27 % of the test prompts that had previously cleared. None of these updates were announced.

Because the rankings reflect live performance rather than launch promises, a tool that held first place in March can fall to fourth by June. The 90-day cadence prevents any single snapshot from dominating the list. It also catches temporary outages and quota experiments that vendors sometimes run for only a few weeks.

We publish the exact retest date for each tool directly beneath its score. If a major model release occurs inside the 90-day window we may accelerate that tool’s retest, but we never shorten the interval for every entry at once. The goal remains consistent: every score you see is no older than three months and is produced by the same documented process.

This fixed rhythm is the only reason the rankings stay current in a category where backend changes happen without notice. It is also why a tool that performs well today can still require verification again soon.