Model research hub

Controlled AI image model comparisons

First-party image model comparisons built from one input, one exact prompt per row, and verified model and output hashes.

Reviewed byRemix Camera Editorial TeamUpdated July 29, 2026

Quick answer

Quick answer

Each current comparison uses one exact input for every model attempt in a row. Prompt hash, input key and hash, profile, aspect ratio, run ID, exact model ID, request fingerprint, and output hash are verified before publication.

At a glance

Benchmark coverage

Current Remix Camera controlled tests
TestModels coveredWhat is controlled
Explicit Seedream benchmarkSeedream 4.5, Seedream V5 Lite, Seedream V5 Pro20 prompts; frozen Mila or Lily input per row; 59 control-valid outputs; one preserved invalid control; one blinded reviewer
Realistic mirror-selfie blind testNano Banana 2, GPT Image 2 Medium, Grok, Seedream V5 Lite, Seedream V5 Pro1 exact prompt; one Lily input; 4 verified outputs; 1 visible block; anonymous reader vote
One prompt across five modelsSeedream 4.5, GPT Image 2 Medium, Grok, Seedream V5 Lite, Seedream V5 Pro5 prompts; one Mila input; 25 verified outputs; blind reader vote; exact prompt, request, model, and output hashes
Adult-fashion moderation boundary testNano Banana 2, GPT Image 2 Medium, Seedream 4.53 prompts; 9 attempts; 7 images; 2 GPT safety blocks; no broader moderation ranking
Seedream version testSeedream 4.5, V5 Lite, V5 Pro5 exact prompts and one input across all three versions
Seedream V5 Lite vs ProV5 Lite and V5 Pro5 priority prompts, including the kitchen portrait, with identical inputs per pair

Curated guides

Start with the guide that matches your goal

Choose well

How to use this collection

  1. Step 1Start with the input shown first in every row.
  2. Step 2Compare outputs only within a row, where the prompt and input are identical.
  3. Step 3Treat the two GPT safety blocks as specific to this three-prompt adult-fashion sample, not a universal policy ranking.
  4. Step 4Choose a winner only after reviewing the full-size outputs for your use case.

How to read these pages

Every row starts with the input. The images that follow are ordered by exact model ID, and the prompt text appears below the grid.

  • Input is byte-identical across every model in a row
  • Prompt text and prompt SHA-256 are identical across the row
  • Profile, aspect ratio, and comparison run ID are fixed
  • Each output is stored under its recorded model ID and output SHA-256

Scope and methodology

These pages replace the older unverifiable rows with a July 2026 controlled rerun. Historical outputs without exact-input proof are excluded from the comparison grids.

Model providers change over time. Treat these as dated results and rerun the same locked method before making a high-volume production decision.

FAQ

Common questions

Which AI image model is best for SFW portraits?

In the July 17 realistic mirror-selfie test, four blinded visual evaluations ranked Seedream V5 Pro first and Grok second. That is a one-prompt result, so compare face fidelity, anatomy, texture, and prompt accuracy for your own use case.

Which AI image model is best for NSFW prompts?

In the July 29 controlled explicit-image benchmark, Seedream V5 Pro led Seedream 4.5 by one decided win across 19 fair three-way rows; V5 Lite trailed. One blinded reviewer scored one output per model per prompt, so treat the result as directional rather than universal.

What should I use when character consistency is critical?

First compare models with the exact same verified input file. For a recurring character across many scenes, a trained LoRA can also help, but its performance should be evaluated in a separate controlled run.

Explore related topic hubs

Ready to try a prompt with your own character?

Browse public prompt packs, choose a look, and remix it with an approved photo or trained character.

Explore prompt packs