NightCafe AI Models Guide

NightCafe AI Models Guide: Choosing the Right Model

Not sure which model to use? Start here

NightCafe offers a lot of models, and that can feel like a lot to take in when you're starting out. The good news is that you don't need to work your way through the list. A small group of models covers almost everything, and these are the ones we'd recommend reaching for first:

If you want…Use thisApprox. credits
One model for almost everything — and the best value in the whole lineupGPT Image 2 Low (PRO)~1
Flagship quality for very few creditsMuse Image or MAI Image 2.6 Flash (PRO)~2 each
More detail and polish from that same modelGPT Image 2 Medium, then GPT Image 2 High~5.5 / ~22
The strongest image editingMAI Image 2.6, GPT Image 2 (any tier), Muse Image~4.25 / ~1+ / ~2
Good results without a PRO subscriptionZ-Image Turbo, or Flux 2 Klein 9B Fast if you also want to edit images~0.5 / ~0.75
Completely free (no credits)Stable Diffusion 1.5 or DreamShaper v80

If you only remember one thing: GPT Image 2 Low is the default recommendation for most people and most prompts. It is the low-quality setting of the highest-ranked image model in the world, and at about one credit per image it is cheaper than most of the older models it replaces.

Think of the models above as the safe choices — the ones most likely to give you a good result whatever you ask them for. The rest of the lineup is well worth exploring once you've found your feet, because every model has its own look, its own quirks and its own strengths. Trying a few is a big part of the fun, and a model that ranks lower on paper is sometimes exactly the aesthetic you were after. The rest of this guide covers those options, and which model to pick when you want something specific.

How to read this guide

  • The model picker is ordered by recommendation. Models near the top of the picker are the ones we suggest first, and this guide follows the same order.
  • Approx. credits is the base cost for one standard image. The final cost changes with resolution, number of images, reference images, and any PRO unlimited allowances.
  • PRO means the model needs a NightCafe PRO subscription. Everything marked otherwise can be used on a free account with credits.
  • Fast credits: a few premium models (GPT Image 2 High, Nano Banana Pro) can only be run with purchased "fast" credits, not earned credits in relax mode.
  • Editing terms: "prompt-to-edit" = change an existing image with a text instruction. "Multi-image fusion" = combine several reference images into one result. "Consistent character" = keep the same character or subject across images (comics, storyboards).

Recommendations below are based on the Artificial Analysis image arenas (blind human preference voting for both text-to-image and image editing) cross-referenced against NightCafe's own credit pricing, as at September 2026.

Best value for money

"Value" means quality per credit, not simply the cheapest option. Because the strongest value picks are PRO models, the lists are split.

Best value — PRO models

ModelApprox. creditsWhy it's good value
GPT Image 2 Low~1The best value model on NightCafe. It is the world's top-ranked image model at its low quality setting, with class-leading prompt accuracy and near-perfect text, plus editing, multi-image fusion and consistent characters. Recommend this first for almost any use-case.
Muse Image~2Meta's flagship. Top-five in the world for text-to-image and top-three for image editing, for two credits. The best quality-per-credit above the sub-1-credit tier.
MAI Image 2.6 Flash~2Microsoft's fast model. Fourth in the world for image editing and top-ten for text-to-image — remarkable at this price.
Qwen Image 2512~1Strong realism and excellent text for a single credit (text-to-image only).
GPT Image 1.5 Low~1The previous OpenAI generation. Still very capable, and supports editing and fusion.
Nano Banana 2 Lite~3.5The cheapest way to get Google's Nano Banana 2 style editing and 1K generation.
Seedream 4.0~3.54K output at a mid-range price.

Best value — no PRO subscription needed

ModelApprox. creditsWhy it's good value
Z-Image Turbo~0.5The best realism per credit without a subscription. Text-to-image only.
Flux 2 Klein 9B Fast~0.75The best all-rounder on a free account: sub-second generation plus prompt-to-edit, multi-image fusion and consistent characters.
Flux Schnell~0.5Very fast and very cheap, at older Flux quality.
Flux 2 Klein 4B Fast~0.5A smaller Klein model — fine for drafts and iteration.
HiDream I1 Fast~1Artistic, high-clarity looks at low cost.
Qwen Image Edit Plus~2The best photo editing available without PRO; also combines reference images.
Flux Kontext Dev~2.5Natural-language photo editing, no PRO required.
Stable Diffusion 1.5, DreamShaper v8FreeFree base generations, at 2022-era quality.

Best model for each task

A fast lookup. Picks are deliberately short — the first model listed is the one to suggest unless the person has a reason to avoid it.

Task / use-caseBest picksWhy
Anything and everythingGPT Image 2 Low, then Medium / High for more detailRanked first in the world for text-to-image and second for editing; cheapest tier costs about 1 credit.
Best value (PRO)GPT Image 2 Low, Muse Image, MAI Image 2.6 FlashFlagship-class quality for 1–2 credits.
Best value (no PRO)Z-Image Turbo, Flux 2 Klein 9B FastStrong results for well under a credit.
FreeStable Diffusion 1.5, DreamShaper v8Free base generations, no PRO required.
Editing an existing imageMAI Image 2.6, GPT Image 2, Muse ImageThe current top three in the image editing arena. Nano Banana 2 Lite (~3.5) is the budget option; Flux 2 Klein 9B Fast is the non-PRO option.
PhotorealismGPT Image 2, MAI Image 2.6, Grok Image 2.0 MediumLifelike people, skin, lighting and natural detail.
Text & typographyGPT Image 2, Nano Banana 2 / Pro, Qwen Image 3.0 ProAccurate spelling and layout for posters, signs and captions.
Complex prompts / prompt adherenceGPT Image 2 High, MAI Image 2.6, Nano Banana ProFollows long, detailed instructions faithfully.
Artistic & stylised workGPT Image 2 Medium, Muse Image, Seedream 4.xPainterly, stylised and concept-art looks.
Logos, graphic design & vectorsRecraft v3, Ideogram V3 Quality, Flux 2 FlexClean design output; Recraft v3 can generate scalable vectors.
Combining reference imagesGPT Image 2, Muse Image, Nano Banana 2 / ProMerge several inputs (e.g. product + scene) into one image.
Consistent characters (comics, storyboards)GPT Image 2, Muse Image, Nano Banana ProKeep the same character across multiple images.
Anime & illustrationAnimagine XL, Blue Pencil XL, AAM XL Anime MixDedicated anime checkpoints; Seedream for modern stylised looks.
Highest resolution (4K)Seedream 4.0 / 4.5, Nano Banana 2Large, print-ready output.
VideoGoogle Veo 3.1, Seedance 2.0, Kling 3.0 TurboText-to-video and image-to-video, several with synchronised audio.
Your own face, object or styleFine-tuned LoRA modelsTrain a model on your own images — see the LoRA section.

These are the models at the top of the model picker, described in more detail. Anything not listed in this section is still available, but there is usually a better or cheaper option here.

OpenAI — GPT Image 2 (and 1.5)

GPT Image 2 is the best model on NightCafe for most things. It is first in the world for text-to-image and second for image editing, with class-leading prompt fidelity, text rendering and layout control, plus prompt-to-edit, multi-image fusion, consistent characters and grounded reasoning. The Low / Medium / High tiers are the same model at different quality settings — they trade credits for detail, not capability.

  • GPT Image 2 Low (~1 credit, PRO) — the default recommendation. Outstanding prompt fidelity and text for a single credit.
  • GPT Image 2 Medium (~5.5 credits, PRO) — step up when Low isn't detailed enough.
  • GPT Image 2 High (~22 credits, PRO, fast credits only) — the highest-fidelity setting, for exceptional detail and text.
  • GPT Image 1.5 (Low / Medium / High) (~1 / 3.5 / 10 credits, PRO) — the previous generation. Still good, and 1.5 Low is a solid one-credit option.
  • GPT Image 1 (Low / Medium / High) — superseded by 1.5 and 2; no longer recommended.

Meta — Muse Image

Muse Image (~2 credits, PRO) is Meta's flagship image model and one of the best value models on NightCafe. It is top-five in the world for text-to-image and top-three for image editing, with prompt-faithful generation, precise edits, multi-reference composition, consistent characters, reasoning and strong text rendering at up to 2K. Recommend it whenever someone wants flagship-class quality but GPT Image 2 Medium feels expensive.

Microsoft — MAI Image 2.6 (new)

Microsoft's MAI Image models are the newest additions to NightCafe, and they arrive at the top of the editing leaderboard.

  • MAI Image 2.6 (~4.25 credits, PRO) — ranked first in the world for image editing and second for text-to-image, behind only GPT Image 2. Versatile quality, strong realism, competitive speed, and prompt-to-edit with up to five reference images.
  • MAI Image 2.6 Flash (~2 credits, PRO) — the fast, affordable variant, and another excellent value pick: fourth in the world for image editing and top-ten for text-to-image.

Google — Nano Banana (Gemini)

Google's Gemini image models (nicknamed "Nano Banana") have first-class start-image support and are excellent for editing and combining photos, with very strong text rendering.

  • Nano Banana 2 Lite (~3.5 credits, PRO) — a faster, cheaper variant for 1K text-to-image and editing.
  • Nano Banana 2 (~7 credits, PRO) — AKA Gemini Flash 3.1. Top-five in the world for both generation and editing, with 4K support, fusion, consistent characters and reasoning.
  • Nano Banana Pro (~11 credits, PRO, fast credits only) — the most capable Nano Banana, with perfect prompt-adherence and typography scores.
  • Nano Banana (Gemini Flash 2.5) (~4 credits, PRO) — retiring on October 2, 2026, and shown with a "Gone Soon" label in the picker. See Deprecation of Nano Banana. Use Nano Banana 2 instead.

xAI — Grok Imagine

  • Grok Image 2.0 Low (~4 credits, PRO) — xAI's latest generation on the faster setting, with 1K/2K output and prompt-to-edit from up to three reference images. Early editing-arena results for Grok 2.0 are very strong.
  • Grok Image 2.0 Medium (~6 credits, PRO) — the default flagship quality for Grok 2.0.
  • Grok Imagine Image Quality (~7 credits, PRO) and Grok Imagine Image (~2 credits, PRO) — the previous generation, still available.

ByteDance — Seedream

Seedream models are fast and high-resolution, and support editing and fusion. They are the go-to for 4K output.

  • Seedream 4.0 (~3.5 credits, PRO) — ultra-fast high-res (2K–4K) generation and editing.
  • Seedream 4.5 (~4.5 credits, PRO) — an improved 4.0, also up to 4K.
  • Seedream 5.0 Pro (~5 credits, PRO) — ByteDance's flagship, with stronger control, multi-image fusion and precise local edits at 2K.
  • Seedream 5.0 Lite (Preview) (~3.5 credits, PRO) — responsive text-to-image with precise prompt adherence.

Black Forest Labs — Flux

The Flux family spans near-free distilled models to high-end production tools, and includes dedicated editing models (Kontext). The Klein models are the reason to come here — they are the best non-PRO options on NightCafe.

  • Flux 2 Klein 9B Fast (~0.75 credits, no PRO) — sub-second generation plus editing, fusion and consistent characters. The recommended free-account model.
  • Flux 2 Klein 9B (~1 credit, PRO) and Klein 4B / 4B Fast (~1 / 0.5 credits, no PRO) — other points on the same speed/quality curve.
  • Flux 2 Max (~7 credits, PRO), Flux 2 Pro (~3), Flux 2 Flex (~6, tuned for typography), Flux 2 Dev (~1.5) — the full Flux 2 line, all with editing, fusion and consistent characters.
  • Flux Kontext (Dev / Pro / Max) (~2.5 / 3.5 / 7 credits) — specialised for editing images with natural-language prompts. Kontext Dev needs no PRO.
  • Flux Schnell (~0.5 credits, no PRO), Flux (~1, PRO), Flux Krea (~1.5, PRO), Flux PRO 1.1 (~4) and 1.1 Ultra (~7, 2K) — the older Flux 1 generation.
  • Juggernaut Flux (Base / Pro) — realism-focused Flux fine-tunes by RunDiffusion.

Alibaba — Qwen Image & Z-Image

  • Z-Image Turbo (~0.5 credits, no PRO) — excellent realism for the price and the best free-account text-to-image model.
  • Qwen Image 2512 (~1 credit, PRO) — natural humans, sharp textures, accurate text for one credit.
  • Qwen Image 3.0 / 3.0 Pro (~3.5 / 7.5 credits, PRO) — the latest Qwen: native 2K, class-leading typography, multi-image editing. 3.0 Pro is top-15 worldwide for both generation and editing.
  • Qwen Image 2.0 / 2.0 Pro (~3.5 / 7.5 credits, PRO) — the previous unified generation-and-editing models.
  • Qwen Image Edit / Edit Plus / Edit 2511 (~1 / 2 / 2 credits) — dedicated editing models. Edit and Edit Plus need no PRO.
  • Qwen Image / Qwen Image SD (~1 / 0.75 credits) — typography-strong base models; SD is cheaper and lower-res.

Other current models

  • HiDream I1 Dev / Full / Fast (~1.5 / 2 / 1 credits) — artistic, high-clarity generation at different speed and cost points. I1 Fast needs no PRO.
  • Ideogram V3 / V3 Quality / V3 Turbo (~4 / 6 / 2 credits, PRO) — strong prompt adherence and text-in-image. Earlier Ideogram 2.x models are also still available.
  • Recraft v3 (~5 credits, PRO) — excellent spelling and typography, and the only model that can output vector graphics.
  • Stable Core (~4 credits, PRO) — beautiful high-resolution images without complex prompts.

Video models

Video models generate short clips from a text prompt and/or a start image, and many include synchronised audio. All video models are PRO, most require a higher PRO plan level, and credit costs are per clip and much higher than image models.

  • Google Veo 3.1 / 3.1 Fast / 3.1 Lite — lifelike motion with context-aware synchronised audio and first/last-frame references. Veo 3.1 is the most popular pick.
  • Seedance 2.0 / 2.0 Fast / 2.0 Mini and Seedance 1.5 Pro — ByteDance's state-of-the-art video with native synchronised audio.
  • Kling 3.0 Turbo and Kling 2.5 Turbo Pro / Standard — cinematic motion, strong prompt adherence and improved lip sync.
  • Wan 3.0 / 3.0 Prime — Alibaba's latest video models.
  • Grok Imagine Video — xAI's multimodal video with native audio from text or a start image.
  • PixVerse V6 / V5 — multi-shot storytelling with synchronised audio and cinematic controls.
  • Happy Horse 1.0 — best-in-class motion quality with synchronised audio.
  • Gemini Omni Flash 1.1 / Omni Flash — Google's multimodal video with native audio.
  • MiniMax H3 and Minimax Hailuo 0.2 — MiniMax's current video models.
  • Runway Gen-4 Turbo, Seedance 1.0 Pro / Pro Fast, Veo 3.0 Fast — earlier models, still available.

LoRAs & fine-tuned (custom) models

In addition to the official models above, you can train your own custom model on NightCafe. These are LoRAs ("Low-Rank Adaptation") — a fast way to fine-tune Stable Diffusion on a small set of images so it can recreate a specific face, object/animal, or style.

How to train a LoRA

  1. Open My Models (in the main menu on desktop, or under your profile picture on mobile) and click Fine-tune a new model.
  2. Choose a model type (face, object/animal, or style) and give it a name.
  3. Choose or create a dataset — upload at least ~20 images. More varied images generally give better results.
  4. Agree to the terms and start training. Training typically takes about 10–30 minutes, and you'll be notified when it's done.

Fine-tuning is a PRO feature, but free users get 1 free face-model tune plus 10 free generations with it. Your trained models are private to your account.

How to use a LoRA in a prompt

To use a fine-tuned model, add its token to your prompt in the format <type:name:optional weight> — for example <lora:My Face:0.8>. You can write prompts like "A photo of <lora:My Face:0.8> riding an elephant" or "A unicorn in the style of <lora:Dark Fantasy:0.5>". The weight is optional (usually between 0 and 1) and defaults to 0.8 if omitted.

Learn more in Introducing Fine-Tuning on NightCafe.

Older model classes (still available)

NightCafe keeps many earlier open-source checkpoints available. They are cheap and can be great for specific styles, but they have much lower prompt adherence and text ability than anything in the recommended list above. Rather than list every checkpoint, they're grouped into classes below.

SDXL models (Stable Diffusion XL)

SDXL launched in 2023 as a major step up from SD 1.5, producing sharper 1024px images. NightCafe hosts many community SDXL checkpoints, each tuned for a look. All are low-cost (~0.5–1 credit), and "Lightning" variants are the fastest and cheapest. Examples by strength:

  • Photorealism: RealVisXL (v3–v5), Crystal Clear XL, Cherry Picker XL, Fluently XL, Juggernaut XL / v9 / XI.
  • Art & concept: DreamShaper XL, Starlight XL, Mysterious XL.
  • Anime & stylised: Animagine XL, Blue Pencil XL, AAM XL Anime Mix, Virtual Utopia XL.
  • Cartoon: Real Cartoon XL.
  • Base & utility: SDXL 1.0, SDXL DPO.

SD 1.5 models (Stable Diffusion 1.5)

SD 1.5 is the first-generation Stable Diffusion model from 2022. It's fast and inexpensive (some are free), but lower resolution and far less accurate than modern models. NightCafe hosts many SD 1.5 community checkpoints, including:

  • Realism: Realistic Vision, AbsoluteReality, Juggernaut Reborn.
  • Art & "does-it-all": DreamShaper v8 (free), NeverEnding Dream.
  • Anime, comics & cartoon: Blue Pencil, Arthemy Comics, 3D Animation Diffusion, RealCartoon Pixar, Ghost Mix.
  • Niche: RPG (character portraits), Rabbit (cute art), WildlifeX (animals), Nightmare Shaper (dark art).

Upscalers

Clarity Upscaler — adds detail and clarity while upscaling up to 4x.

Original algorithms

NightCafe's earliest algorithms remain available for nostalgia and unique looks: Artistic (VQGAN+CLIP, the original text-to-image method), Coherent (CLIP-guided diffusion), and Style Transfer (neural style transfer).

Retired / deprecated models

Some models have been retired, usually because an external provider discontinued them. These are no longer available for new creations (or are scheduled to be removed on the dates below):

Understanding model attributes

NightCafe rates most image models on four qualities, each scored from 0 to 1 (higher is better). These power the "best for" picks above:

  • Prompt adherence — how faithfully the model follows your prompt.
  • Art — strength at artistic and stylised output.
  • Realism — photorealistic quality.
  • Typography — ability to render readable, correct text.

Base cost is the approximate credits for one standard image. Actual credits depend on resolution, number of images, start/reference images and your plan. See the credits and subscription articles for details.

FAQ

There are too many models — which one should I use? GPT Image 2 Low. It's the cheapest tier of the world's highest-ranked image model, at around 1 credit an image, and it handles generation, editing and reference images. Move up to GPT Image 2 Medium if you want more detail.

Which models give the most for my credits? On PRO: GPT Image 2 Low (~1), Muse Image (~2) and MAI Image 2.6 Flash (~2). Without PRO: Z-Image Turbo (~0.5) and Flux 2 Klein 9B Fast (~0.75).

Which models are free? Stable Diffusion 1.5 and DreamShaper v8 offer free base generations. Most modern models require credits and many require PRO.

Which models can edit my photos? MAI Image 2.6, GPT Image 2 and Muse Image are the strongest. Nano Banana 2 Lite is the cheap PRO option, and Flux 2 Klein 9B Fast or Qwen Image Edit Plus work without PRO.

Which models make video? See the Video models section — Veo 3.1, Seedance, Kling, Wan 3.0, Grok Imagine Video and PixVerse are current options.

Can I make my own model? Yes — train a LoRA on your own face, object or style. See the LoRA section above.

Model availability and pricing change over time as providers release new models and retire old ones. This guide reflects NightCafe's official model lineup and the Artificial Analysis image arena rankings at the time of writing. (7th September 2026)