The AI image generator evaluation checklist.
Fifteen questions, organised in five sections. Print it. Save it. Reuse it on every new generator that launches. The checklist is updated when the category evolves; this version is current for April 2026.
How to use this checklist.
For each generator you are evaluating, walk through the fifteen questions in order. Most can be answered in 30-60 minutes from vendor documentation; a few (test protocols on /capabilities) take longer. Score each answer as Confirmed, Partial, or Unconfirmed. A generator with all 15 confirmed is rare; a generator with most confirmed and a known gap is fine, as long as the gap is acceptable for your use case. Beware generators where multiple answers are unconfirmed because the docs are silent.
Capability.
- 01
Can it render the resolution I need natively? What is the upscaling workflow?
Why this matters · How to find the answer
Why: Upscaling can introduce artefacts and hallucinations. Native resolution determines how much detail the model put in versus interpolated.
Find it: Vendor docs, "model specifications" or "output resolution" section. Look for the native resolution and the upscaling pipeline (latent, ESRGAN, separate model).
☐ Confirmed☐ Partial☐ Unconfirmed - 02
How does it handle text in images for my prompt style?
Why this matters · How to find the answer
Why: Text rendering remains uneven across generators. If your use case has any text-on-image requirement, this is a top-three axis.
Find it: Run the test prompts on /capabilities#text-rendering. Also check whether the vendor publishes the text-encoder size (T5-XXL is a positive signal).
☐ Confirmed☐ Partial☐ Unconfirmed - 03
Can it produce consistent characters or subjects across multiple images?
Why this matters · How to find the answer
Why: Subject consistency is architecturally difficult. The mechanisms (LoRAs, reference images, character sheets) are not equally available across vendors.
Find it: Look for "character consistency", "reference image", "character sheet" in vendor docs. LoRA support requires open weights or a vendor-side fine-tuning facility.
☐ Confirmed☐ Partial☐ Unconfirmed - 04
Does it support reference images or style transfer?
Why this matters · How to find the answer
Why: Reference images let you pin output to a defined style. Without them, prompt engineering bears more weight.
Find it: Check for "image-to-image", "reference", "style transfer", "IP-Adapter" in the docs.
☐ Confirmed☐ Partial☐ Unconfirmed - 05
How does it handle negative prompts?
Why this matters · How to find the answer
Why: Negation is hard for diffusion; some generators support it natively, others require workaround techniques.
Find it: Vendor docs "prompt syntax" or "advanced parameters" sections.
☐ Confirmed☐ Partial☐ Unconfirmed
Licensing.
- 06
Does my plan permit commercial use? What clause confirms this?
Why this matters · How to find the answer
Why: Free and personal-tier plans often restrict commercial use, sometimes with carve-outs that are easy to miss.
Find it: Plan-tier comparison on the pricing page. Plus the "Content" or "Output" section in the terms of service.
☐ Confirmed☐ Partial☐ Unconfirmed - 07
Does the vendor offer indemnification at my plan tier?
Why this matters · How to find the answer
Why: Indemnification protects you against IP claims arising from outputs. Most consumer plans do not include it.
Find it: Enterprise contract page or "Copyright Shield" / "IP indemnification" section in the terms.
☐ Confirmed☐ Partial☐ Unconfirmed - 08
What is the output copyright status in my jurisdiction based on the vendor's docs?
Why this matters · How to find the answer
Why: Whether you can register the output as your copyright varies by jurisdiction. Vendor docs sometimes address this.
Find it: Vendor FAQ on copyright. Cross-reference with /licensing#copyright on this site.
☐ Confirmed☐ Partial☐ Unconfirmed - 09
Can I opt my generated content out of being used as training data?
Why this matters · How to find the answer
Why: Free tiers often grant the vendor rights to your prompts and outputs as training data. Opt-out is sometimes available, sometimes plan-tier-gated.
Find it: "Data use" or "Feedback" section in the terms. Sometimes a separate privacy or training-opt-out page.
☐ Confirmed☐ Partial☐ Unconfirmed
Provenance.
- 10
What is the training data source, and is it disclosed?
Why this matters · How to find the answer
Why: The legal posture of the model traces to its training data. Disclosed sourcing signals confidence in the licensing position.
Find it: "About the model", "Training", or "Data" sections in vendor docs. Look for explicit sourcing rather than "carefully curated".
☐ Confirmed☐ Partial☐ Unconfirmed - 11
Can rights-holders opt out, and has the vendor honoured opt-outs?
Why this matters · How to find the answer
Why: If you are a rights holder yourself, opt-out matters directly. If you are a buyer, public commitment to opt-outs signals the vendor's overall provenance posture.
Find it: "Opt-out", "Spawning", "HaveIBeenTrained", "robots.txt", "TDM reservation" in vendor pages.
☐ Confirmed☐ Partial☐ Unconfirmed
Cost and integration.
- 12
What is the per-image cost at my expected volume?
Why this matters · How to find the answer
Why: Pricing structures vary (per-image, per-megapixel, per-step, monthly cap). The right comparison depends on your specific volume profile.
Find it: Pricing page. For API products, the rate card. Calculate your per-image cost using your expected resolution and steps.
☐ Confirmed☐ Partial☐ Unconfirmed - 13
Does the API rate limit, latency, and batch support match my use case?
Why this matters · How to find the answer
Why: Production deployment hits these constraints. A great model with a 5-request-per-minute rate limit fails at scale.
Find it: API docs "rate limits" section. Check p50 and p99 latency; check whether batch endpoints exist.
☐ Confirmed☐ Partial☐ Unconfirmed
Workflow and exit.
- 14
Can I export raw prompts, seeds, and generation metadata for reproducibility?
Why this matters · How to find the answer
Why: Reproducibility matters for design workflows where you iterate on a successful generation.
Find it: Vendor UI export options. Some platforms strip metadata on export to non-paid tiers.
☐ Confirmed☐ Partial☐ Unconfirmed - 15
If I cancel, can I retain the images? What becomes of them on the vendor side?
Why this matters · How to find the answer
Why: Lock-in risk. Some platforms delete free-tier outputs after a period or on cancellation.
Find it: "Account closure" or "Data retention" in the terms. Confirm whether download is possible at any time.
☐ Confirmed☐ Partial☐ Unconfirmed