Text-to-image models

Text-to-image models: prices, adult policies and bias findings

The catalog's portraits are generated with text-to-image models, and this page lists every one examined for it: the 22 models below, grouped by whether the site calls them, with each fact given its source and the date it was read. Prices are each provider's published rate on that date. Nothing here ranks the models, and nothing here is a number about a people.

A side-by-side comparison of the generators the site uses, scored against documented photographs and judged by readers, is being built and is not yet published. When it is, it appears here and each model in the comparison links to its own page.

In the catalog pipeline today

The models the site itself calls to generate portraits. Recraft V3 is the default; Flux 1.1 Pro is what the member generator pins; the Stability models are the fallback through AWS Bedrock.

  • Recraft V3Recraft

    Called through
    Replicate: recraft-ai/recraft-v3
    Price
    $0.04 per image (source, read 2026-09-07)
    Adult content
    Returns a clothed image on an explicit prompt, with no error, and the run is charged. Measured on this site.
    Negative prompt
    Not accepted on Replicate; accepted on Recraft’s own API
    Seed
    Not accepted on Replicate; Recraft’s own API has random_seed
    Output terms
    Replicate: outputs may be used commercially. Recraft: the paid tier assigns copyright in outputs to the customer, the free tier does not; the terms bar building a competing service. (terms, read 2026-09-07)
    Bias findings
    • On this site, held its phenotype on Aboriginal Australian, Norwegian, Japanese and Yoruba in the 2026-05-17 head-to-head that made it the catalog default. No published audit found. (Site log, 2026-05-17)

  • FLUX 1.1 ProBlack Forest Labs

    Called through
    Replicate, or the BFL API: black-forest-labs/flux-1.1-pro
    Price
    $0.04 per image (source, read 2026-09-07)
    Adult content
    Renders an explicit prompt at safety_tolerance 5 (the Replicate scale runs 1 to 6; the benchmark sends the default). Measured on this site.
    Negative prompt
    Not accepted
    Seed
    Accepted
    Output terms
    BFL: the customer owns all right, title and interest in output, commercial use allowed; the API service terms say BFL may use inputs and outputs to train; the usage policy bans sexual depiction of real people and any minor and states no ban on fictional adult content. (terms, read 2026-09-07)
    Bias findings
    • Collapsed Aboriginal Australian portraits to an East African phenotype in the 2026-05-17 head-to-head on this site. (Site log, 2026-05-17)
    • Described as weak on rendering ethnicity in a user discussion (August 2024). (github.com/lllyasviel/Fooocus/discussions/3561)
    • FLUX 1.0-dev scored the lowest alignment of four models in CulturalFrames (0.63). (Nayak et al. 2025, arXiv 2506.08835)
    • FLUX-Dev lost more identity on dark-skinned faces in a large facial-editing evaluation. (Nair, Pandit, Kerner 2026, arXiv 2608.28802)

  • FLUX.1 schnellBlack Forest Labs

    Called through
    Replicate: black-forest-labs/flux-schnell
    Price
    $0.003 per image (source, read 2026-09-07)
    Adult content
    Returns no image on an explicit prompt; the run fails and is not charged. Measured on this site.
    Negative prompt
    Not accepted
    Seed
    Accepted; not deterministic when go_fast is on
    Output terms
    Apache 2.0 weights; Replicate: the customer owns outputs. (terms, read 2026-09-07)
    Bias findings
    • Included in the INFELM skin-tone alignment benchmark (2025); per-model figure not extracted here. (Jin et al. 2025, arXiv 2501.01973)

  • Stable Image CoreStability AI

    Called through
    AWS Bedrock: stability.stable-image-core-v1:1
    Price
    $0.04 per image (source, read 2026-09-07)
    Adult content
    Returns a clothed image on an explicit prompt, as Recraft does. Stability’s use policy also forbids content relating to sexual intercourse or sexual acts. Measured on this site.
    Negative prompt
    TBD on Bedrock
    Seed
    TBD on Bedrock
    Output terms
    Stability EULA and the AWS Responsible AI policy. (terms, read 2026-09-07)
    Bias findings
    • TBD (no audit found)

  • Stable Image UltraStability AI

    Called through
    AWS Bedrock: stability.stable-image-ultra-v1:1
    Price
    $0.14 per image (source, read 2026-09-07)
    Adult content
    Sanitizes like Stable Image Core; the same use policy applies. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    TBD on Bedrock
    Seed
    TBD on Bedrock
    Output terms
    Stability EULA and the AWS Responsible AI policy. (terms, read 2026-09-07)
    Bias findings
    • TBD (no audit found)

  • Stable Diffusion 3.5 LargeStability AI

    Called through
    AWS Bedrock: stability.sd3-5-large-v1:0
    Price
    $0.08 per image (source, read 2026-09-07)
    Adult content
    Returns filter reason codes rather than a silent substitute; the same use policy applies. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Accepted, up to 10,000 characters
    Seed
    Accepted
    Output terms
    Stability EULA and the AWS Responsible AI policy. (terms, read 2026-09-07)
    Bias findings
    • SD 3.5 Large scored 0.66 alignment in CulturalFrames; the Stable Diffusion family is the most audited lineage in the bias literature. (Nayak et al. 2025, arXiv 2506.08835)

Candidates not used on this site

Reachable by API and priced, examined for the comparison, not called by the catalog.

  • Ideogram 3.0 (turbo, balanced, quality)Ideogram

    Called through
    Replicate, or Ideogram’s own API: ideogram-ai/ideogram-v3-turbo, ideogram-ai/ideogram-v3-balanced, ideogram-ai/ideogram-v3-quality
    Price
    $0.03, $0.06 and $0.09 per image for turbo, balanced and quality (source, read 2026-09-07)
    Adult content
    Ideogram’s API terms forbid obscene, lewd, lascivious, filthy or pornographic content. Behavior not measured. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Not accepted
    Seed
    Accepted
    Output terms
    Ownership of outputs is not stated explicitly in the API terms; no benchmark clause found. (terms, read 2026-09-07)
    Bias findings
    • TBD (no audit found)

  • Gemini 2.5 Flash Image and newer (Nano Banana)Google

    Called through
    Replicate, or the Gemini API: google/nano-banana, google/nano-banana-2, google/nano-banana-pro, gemini-2.5-flash-image
    Price
    $0.039 per image on the Gemini API; Nano Banana 2 $0.067; Replicate price TBD (source, read 2026-09-07)
    Adult content
    Google’s prohibited-use policy forbids content created for the purpose of pornography or sexual gratification; every output carries a SynthID watermark. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Not accepted
    Seed
    Not accepted on Replicate
    Output terms
    TBD
    Bias findings
    • On neutral prompts, 96.4% of faces across Gemini 2.5 Flash Image and GPT Image 1.5 were classified white; Gemini’s were 93.7% female with mean Monk skin tone 4.2 to 5.1. (Balestri 2026, arXiv 2602.12133)
    • Its predecessor Imagen 3 placed 59% of depicted people at Monk 1 to 3 and 0.3% at 9 to 10, by Google’s own report. (Google DeepMind 2024, arXiv 2408.07009)

  • gpt-image-1, 1.5 and 2OpenAI

    Called through
    OpenAI direct, or Replicate for 1.5 and 2: gpt-image-1, openai/gpt-image-1.5, openai/gpt-image-2
    Price
    gpt-image-1 at 1024 by 1024: $0.011 low, $0.042 medium, $0.167 high quality; Replicate prices TBD (source, read 2026-09-07)
    Adult content
    Returns a moderation_blocked error with the category sexual rather than a substitute image; a lower moderation setting exists. The usage policy page refused the fetch. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Not accepted
    Seed
    Not accepted
    Output terms
    TBD (the usage policy page returned 403 on the read date)
    Bias findings
    • GPT-Image had the best alignment of four models in CulturalFrames (0.85). (Nayak et al. 2025, arXiv 2506.08835)
    • GPT Image 1.5 rendered 70.6% male on neutral prompts with mean Monk skin tone 3.4 to 3.9. (Balestri 2026, arXiv 2602.12133)
    • Its predecessor DALL-E depicted hospital C-suite images as White in 94.5% of cases. (Gisselbaek et al. 2025, PMC12140938)

  • Seedream 4ByteDance

    Called through
    Replicate, or BytePlus ModelArk: bytedance/seedream-4
    Price
    $0.03 per image on BytePlus, charged per successful image; Replicate price TBD (source, read 2026-09-07)
    Adult content
    TBD. A third party reports Seedream 5 Lite rewriting prompts to a fully dressed subject; not measured here.
    Negative prompt
    Not accepted
    Seed
    Not accepted on Replicate
    Output terms
    TBD
    Bias findings
    • TBD (no audit found)

  • HiDream-I1 fullHiDream

    Called through
    Replicate: prunaai/hidream-l1-full
    Price
    about $0.028 per run, billed per second of GPU time (source, read 2026-09-07)
    Adult content
    MIT license with no content clause; Replicate’s acceptable use policy applies. Not measured. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Not accepted
    Seed
    Accepted
    Output terms
    MIT license. (terms, read 2026-09-07)
    Bias findings
    • TBD (no audit found)

  • Qwen-ImageAlibaba

    Called through
    Replicate: qwen/qwen-image
    Price
    TBD on Replicate
    Adult content
    Apache 2.0 weights with a disable_safety_checker flag on Replicate; Replicate’s acceptable use policy applies. Not measured. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Accepted
    Seed
    Accepted
    Output terms
    Apache 2.0. (terms, read 2026-09-07)
    Bias findings
    • Included in Culture in Action (2025), where Global North prompts scored 4 to 8 points higher across all six models tested. (Malakouti, Gong, Kovashka 2025, arXiv 2511.05681)

  • Stable Diffusion 3.5 Large and MediumStability AI

    Called through
    Replicate: stability-ai/stable-diffusion-3.5-large, stability-ai/stable-diffusion-3.5-medium
    Price
    $0.065 per image (Large), $0.035 (Medium) (source, read 2026-09-07)
    Adult content
    Stability’s use policy forbids content relating to sexual intercourse or sexual acts. Not measured on this path. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Accepted
    Seed
    Accepted
    Output terms
    Stability Community License: free under $1M annual revenue; the customer owns generated images. (terms, read 2026-09-07)
    Bias findings
    • SD 3.5 Large scored 0.66 alignment in CulturalFrames. (Nayak et al. 2025, arXiv 2506.08835)

Adult-capable open weights, not used on this site

Community checkpoints hosted on Replicate that render nudity. Listed for completeness; the catalog sends clothed documentary prompts only and none of these has been called.

  • Pony Realism v2.2community checkpoint

    Called through
    Replicate: aisha-ai-official/pony-realism-v2.2
    Price
    about $0.0067 per run, billed per second of GPU time (source, read 2026-09-07)
    Adult content
    An adult-capable checkpoint; no safety checker documented. Not called by this site. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Accepted
    Seed
    Accepted
    Output terms
    License TBD.
    Bias findings
    • TBD (no audit found)

  • Juggernaut XL v7 and LightningRunDiffusion

    Called through
    Replicate: asiryan/juggernaut-xl-v7, fofr/juggernaut-xl-lightning
    Price
    about $0.0064 and $0.0054 per run, billed per second of GPU time (source, read 2026-09-07)
    Adult content
    RunDiffusion states the NSFW build needs a license for commercial use. Not called by this site. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Accepted
    Seed
    Accepted
    Output terms
    Commercial license TBD.
    Bias findings
    • TBD (no audit found)

  • RealVisXL V3community checkpoint

    Called through
    Replicate: sdxl-based/realvisxl-v3
    Price
    about $0.0034 per run, billed per second of GPU time (source, read 2026-09-07)
    Adult content
    Its model card states it can produce SFW and NSFW images; Replicate’s safety checker can be disabled through the API. Not called by this site. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Accepted
    Seed
    Accepted
    Output terms
    CreativeML Open RAIL++-M. (terms, read 2026-09-07)
    Bias findings
    • TBD (no audit found)

  • NSFW Flux devcommunity checkpoint on FLUX.1 dev

    Called through
    Replicate: aisha-ai-official/nsfw-flux-dev
    Price
    about $0.013 per run, billed per second of GPU time (source, read 2026-09-07)
    Adult content
    Built for adult output; Replicate’s acceptable use policy applies. Not called by this site. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Not accepted
    Seed
    Accepted
    Output terms
    FLUX.1 dev weights are non-commercial off Replicate; Replicate states images generated on Replicate with FLUX.1 models may be used commercially. License of the checkpoint itself TBD. (terms, read 2026-09-07)
    Bias findings
    • TBD (no audit found)

Examined and excluded

Each with the reason it is out.

  • Luma Photon and Photon FlashLuma

    Excluded: Excluded because the terms forbid publishing benchmarks: the prohibited-uses list in lumalabs.ai/legal/tos reads, verbatim, "(g) publish benchmarks or performance information about the Services" (read 2026-09-08).

    Called through
    Luma API, or Replicate: luma/photon, luma/photon-flash
    Price
    $0.015 per 1080p image, Flash $0.002 (source, read 2026-09-07)
    Adult content
    The terms of service forbid generating pornographic or sexually explicit content. Provider policy. (policy, read 2026-09-08)
    Negative prompt
    Not accepted
    Seed
    Accepted
    Output terms
    The customer owns outputs; commercial use only under a paid subscription. (terms, read 2026-09-07)
    Bias findings
    • TBD (no audit found)

  • MidjourneyMidjourney

    Excluded: No official API exists, and its terms bar automated generation.

    Called through
    No official API
    Price
    Not applicable
    Adult content
    Not applicable
    Negative prompt
    Not applicable
    Seed
    Not applicable
    Output terms
    Not applicable
    Bias findings
    • In 100 Midjourney images for "an Indian person", 92 of the subjects wore a traditional pagri or similar headwear; 99 of the 100 images for "a Mexican person" featured a sombrero or similar hat. (Rest of World, 2023-10-10, restofworld.org/2023/ai-image-stereotypes)
    • Midjourney overrepresented White individuals at 75.0% of hospital C-suite images (DALL-E 94.5%, Imagen 51.5%). (Gisselbaek et al. 2025, PMC12140938)

  • Imagen 4Google

    Excluded: The Gemini API documentation states that Imagen is deprecated and will shut down on August 17, 2026. The Gemini image models replace it in the candidate list.

    Called through
    Replicate, or Vertex AI: google/imagen-4, imagen-4.0-generate-001
    Price
    TBD
    Adult content
    Google’s prohibited-use policy forbids sexually explicit content. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Not supported on Imagen 3.0 and newer
    Seed
    Accepted on Vertex with the watermark off
    Output terms
    TBD
    Bias findings
    • Imagen 3 placed 59% of depicted people at Monk 1 to 3 and 0.3% at 9 to 10, by Google’s own report. (Google DeepMind 2024, arXiv 2408.07009)

  • Amazon Nova CanvasAmazon

    Excluded: Bedrock lists the model as Legacy with an end-of-life date of September 30, 2026.

    Called through
    AWS Bedrock: amazon.nova-canvas-v1:0
    Price
    TBD
    Adult content
    The AWS Responsible AI policy applies; individual images can be blocked after generation. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    Accepted, 1 to 1,024 characters
    Seed
    Accepted
    Output terms
    TBD
    Bias findings
    • TBD (no audit found)

  • Leonardo PhoenixLeonardo.Ai

    Excluded: Its pricing and terms pages refused the read, so nothing about it can be stated with a source.

    Called through
    Leonardo production API
    Price
    TBD (pricing pages refused the fetch)
    Adult content
    The API blocks NSFW prompts by default with a 400 error, and flags outputs it judges NSFW. Provider policy. (policy, read 2026-09-07)
    Negative prompt
    TBD
    Seed
    TBD
    Output terms
    TBD (terms page refused the fetch)
    Bias findings
    • TBD (no audit found)