66 MODELS · 12 PROVIDERS · 4 MODALITIES

Generate with every model.Keep every result.

Made here · From the public gallery · Updated continuously

OpenAIGoogle Geminifal.aiNovelAIByteDance SeedreamElevenLabsFish AudioReplicateHyper3D RodinRunwayBlack Forest LabsHugging Face
SCROLLSWIPE

01 · IMAGE

One line. Four models answer.

Image workbenchGPT Image 2Gemini 3 ProFLUX 2 ProSeedream 5.01024×1365 · 1 each · same prompt

GPT Image 2 · Ginger cat curator in round glasses at a museum desk, framed art behind, warm lightGPT Image 2
Gemini 3 Pro · Ginger cat curator in round glasses at a museum desk, framed art behind, warm lightGemini 3 Pro
FLUX 2 Pro · Ginger cat curator in round glasses at a museum desk, framed art behind, warm lightFLUX 2 Pro
Seedream 5.0 · Ginger cat curator in round glasses at a museum desk, framed art behind, warm lightSeedream 5.0
7 models →

02 · LORA

Find one, load it. Load as many as you like.

LoRA workbench0 loaded · no cap
⌕ Search Civitai / HuggingFace…
Flatline
STYLE · AnimaLoad
Anime Figure
STYLE · AnimaLoad
Watercolor
STYLE · SDXLLoad
Dynamic Poses
POSE · AnimaLoad
Detail Tweaker
DETAIL · AnimaLoad
Light Concepts
LIGHT · AnimaLoad
Loaded · weights
Flatline2.0
Anime Figure0.1
Dynamic Poses0.3
Detail Tweaker0.6
Light Concepts0.4
PROMPT · trigger words arrive on their own
@flatlinef1gur3dispersion, full body, standing, cream coat, holding a black book
One mount list · one seed · only the style weights move
The same character at flat-line weight 2 and figure weight 0.12.0 / 0.1
The same character at flat-line weight 1.4 and figure weight 0.71.4 / 0.7
The same character at flat-line weight 0.7 and figure weight 1.40.7 / 1.4
The same character at flat-line weight 0.1 and figure weight 20.1 / 2.0

03 · AUDIO

Pick a voice, type, and it speaks.

Voice room8 speakers · voice library + cloning · Fish S2.1
QQing · Fish soft femaleThe show opens tonight at eight.0:02
LLei · Fish young maleGot it. I'll do one more pass on the poster.0:03
NNarrator · cloned voiceAnd just then, the lamps in the hall came on one by one.0:04
QLN＋Type a line, pick a voice, hit generate…

04 · VIDEO

Image, voice, footage, text — all of it can be a reference.

Video workbenchSeedance 2.5 reference endpoint · 30s in one pass · 720p
▶ shot01-entrancematch this shot
character-anchor-her.webpmatch this character
♪line-her-umbrella 0:03match this voice
＋⬡▤✥4 references · any type · voice cloning ×3
Seedance · made here · The Borrowed Umbrella

05 · CANVAS

Talk out a script, grow it into nodes, wire it into a cut.

1Assistantscript brain
MeRainy season, a girl forgets her umbrella — turn it into a short.
AIThree shots: waiting at the entrance, sharing the umbrella, parting at the arcade. Here is the script.
Script · The Borrowed Umbrella v1
→
2Script3 shots · The Borrowed Umbrella
Shot 01At the entrance, she watches the rain with no umbrella→ node
Shot 02He tilts the umbrella her way, his right shoulder wet→ node
Shot 03Under the arcade, the rain suddenly goes quiet→ node
→
3Canvas4 nodes · 3 wires
Shot 01 · entrance
Shot 02 · shared umbrella
Shot 03 · arcade
Cut · SEEDANCE · made here

06 · VAULT

Everything you keep can go back on stage.

VaultImages · video · audio · LoRA · kept for good
AllImagesVideoAudioLoRA
▶ 05 · The Borrowed Umbrella
04 · character anchor
03 · line 0:03
Image
Image
LoRA · Ghibli Soft
Rainy season, walking home — he tilts the umbrella her way, his right shoulder soaked.PROMPT
3D · RODIN
Image
+ 1,284still in the vault
REUSE
Reference slot · emptyReference loaded · canvas

Any one of these, dragged into the canvas or a workbench, is the reference for the next generation. A result is not the end, it is material.

Start creating →
GPT Image 2 sample outputMade here · GPT Image 2
IMAGE · OPENAI

GPT Image 2

The all-rounder: top of the blind-test boards, and the best at reading a complicated brief.

Poster layoutLong briefsMultilingual type
Usage-basedOn ANTEI · 1 route
Strong at
  • Text inside the image lands almost every time — poster-grade
  • Highest fidelity on long multi-condition prompts; nothing gets dropped
  • Chinese, Japanese and Korean render just as cleanly
Weak at
  • Thinking mode adds latency; it is on the slow side
  • Expensive in bulk, and transparency support is only fair
Gemini 3 Pro Image sample outputMade here · Gemini 3 Pro Image
IMAGE · GOOGLE

Gemini 3 Pro Image

Fast and steady workhorse: 2–5 seconds an image, and the finest editing control here.

Fast iterationLocal editsReal places
$0.13 / image · Flash tier $0.03–0.07On ANTEI · 3 routes
Strong at
  • 2–5 seconds an image — editing never breaks your rhythm
  • Finest control over inpainting, relighting and camera moves
  • Search-grounded: real landmarks and real light have a source
Weak at
  • Holds back on the surreal and the heavily stylised
  • Small faces and small type still break
FLUX 2 Pro sample outputfal.ai official demo output
FLUXIMAGE · BLACK FOREST LABS

FLUX 2 Pro

The texture one: skin and cloth without the plastic sheen, and rock-steady on long descriptions.

Product textureConcept artValue
$0.03 / image · edit $0.04 · Flash $0.005On ANTEI · 5 routes
Strong at
  • Real texture: little of the plastic sheen, sharp on objects and materials
  • Dual-stream architecture holds long descriptive prompts precisely
  • Best value at this quality tier, with no parameter fiddling
Weak at
  • Portraits drift towards a porcelain-doll over-retouch
  • Text inside the image is still unreliable
Seedream 5.0 sample outputMade here · Seedream 5.0
ByteDanceSeedIMAGE · BYTEDANCE

Seedream 5.0

A designer's head: it understands layout and Chinese, and searches the web while it generates.

Chinese postersInfographicsLayout
$0.03–0.09 / image · three integrationsOn ANTEI · 5 routes
Strong at
  • Design-grade sense of layout: a magazine cover in one pass
  • Native type in 10+ languages, strongest in Chinese
  • Searches the web mid-generation, so the facts have a source
Weak at
  • Long strings of text still come out wrong, and it takes 8–15 seconds
  • Photographic texture reads AI-ish; weaker than Gemini at realism
Recraft V4 Pro sample outputMade here · Recraft V4 Pro
RECRAFTIMAGE · RECRAFT

Recraft V4 Pro

The designer's tier: the only one that outputs native SVG, and it sets whole paragraphs.

Vector logosBrand collateralFull paragraphs
$0.21 / imageOn ANTEI · 1 route
Strong at
  • The only model that generates true vector SVG
  • Readable paragraphs where others manage 2–3 words
  • Top of the HF text-to-image arena since October 2024
Weak at
  • $0.21 an image is the most expensive in the image section
  • Design-specialised: general photorealism is not its ground
NovelAI Diffusion V5 sample outputNovelAI Diffusion V5 sample outputNovelAI Diffusion V5 sample outputMade here · NovelAI Diffusion V5
NOVELAIIMAGE · ANLATAN

NovelAI Diffusion V5

The newest anime specialist: 22 characters in one frame without blending, and a whole comic page in one pass.

Anime figuresMany charactersComic panels
Usage-based · V4.5 tier from $0.01 / imageOn ANTEI · 4 routes (V5 ×2 + V4.5 ×2)
Strong at
  • Twice the size plus a 32-channel VAE: detail down to hair strands and trinkets
  • Up to 22 characters in one frame — multi-panel comics lay out directly
  • Prompts in Japanese, traditional Chinese and more; transparency supported
Weak at
  • Anime-specialised; not for realistic scenes
  • Not in the unlimited tier, and billed tighter than V4.5
Illustrious XL sample outputIllustrious XL sample outputIllustrious XL sample outputMade here · Illustrious XL
ILLUSTRIOUSIMAGE · ONOMA AI · 开源

Illustrious XL

The leading open anime base: enormous character knowledge and the busiest fine-tuning scene.

Anime charactersFan workLoRA base
$0.14 / image (direct API)On ANTEI · 1 API route + its own entry in LoRA
Strong at
  • Trained on all of Danbooru: the widest character recall here
  • Far steadier structure — hands and feet — than its open-source generation
  • The busiest fine-tune scene (WAI and others); you will meet it again in LoRA
Weak at
  • Character knowledge stops mid-2024; it will not know a new season
  • Realism and design work are not its race
1 / 7 · GPT Image 2
Illustrious XL sample outputIn-app base-model cover
ILLUSTRIOUSLORA · ONOMA AI · 开源

Illustrious XL

The base-model entry: the anime foundation with the liveliest LoRA scene.

Anime baseLiveliest sceneTwo routes
$0.14 / image (direct API)On ANTEI · API and Runner, both
Strong at
  • The largest Civitai scene: the fullest set of character and style LoRAs
  • Full Danbooru grounding — character terms land straight away
  • The API route needs no deployment; the Runner route mounts anything
Weak at
  • Weak towards realism
  • New-season characters have to come from a LoRA
WAI-Illustrious sample outputIn-app base-model cover
WAILORA · RUNNER

WAI-Illustrious

The crowd-favourite tune: a polished anime face straight out of the box.

Community favouriteNo setupAnime characters
Usage-based (Runner)On ANTEI · self-hosted Runner
Strong at
  • A fixture on the community charts — good-looking on default settings
  • Takes Illustrious-family LoRAs, so characters reproduce reliably
  • Runs straight on the Runner with every parameter open
Weak at
  • Leans sweet; not the one for hard realism
  • Queue depends on Runner load
Anima Pencil-XL sample outputIn-app base-model cover
PENCIL-XLLORA · RUNNER

Anima Pencil-XL

Pencil and clean line: sketch-feel illustration in one pass.

Clean lineSketch feelIllustration
Usage-based (Runner)On ANTEI · self-hosted Runner
Strong at
  • A pencil and clean-line texture nothing else here has
  • Works for line art and for finished colour alike
  • Runs straight on the Runner with every parameter open
Weak at
  • Narrow subject range: illustration is the point
  • Not made for thick-paint realism
Pony Diffusion V6 XL sample outputIn-app base-model cover
PONY V6LORA · RUNNER

Pony Diffusion V6 XL

The grandfather of the style universe: somebody has trained every look on it.

Style sceneThe elderBroad support
Usage-based (Runner)On ANTEI · self-hosted Runner
Strong at
  • The widest LoRA compatibility of any base here
  • A mature system of style trigger words
  • The deepest community back catalogue
Weak at
  • Plain faces out of the box; LoRAs do the lifting
  • Prompt syntax has a learning curve (the score terms)
SDXL 1.0 sample outputIn-app base-model cover
SDXLLORA · RUNNER

SDXL 1.0

The general-purpose veteran: realism through illustration, it takes it all.

General purposeVeteranRealism-capable
Usage-based (Runner)On ANTEI · self-hosted Runner
Strong at
  • The open base with the broadest ecosystem coverage
  • Realism, design and illustration all run on it
  • The most mature tooling and the most documentation
Weak at
  • Raw quality has been left behind by newer models
  • Standing out takes a LoRA stack
Anima（DiT） sample outputIn-app base-model cover
ANIMALORA · RUNNER

Anima（DiT）

The new-architecture testbed: an anime newcomer on a DiT foundation.

DiT architectureNewcomerTestbed
Usage-based (Runner)On ANTEI · self-hosted Runner (v4 pipeline)
Strong at
  • A new DiT architecture with a high ceiling for detail
  • The v4 pipeline is already deployed here
  • A fresh option on the anime side
Weak at
  • The scene is only starting; few LoRAs
  • Generation takes longer
1 / 6 · Illustrious XL
Made here · Seedance 2.0 Fast
ByteDanceSeedVIDEO · BYTEDANCE

Seedance

Reference generation in one go: the video workhorse, on 18 routes.

Reference inputMany routesVoice cloning
$0.12–0.30 / secondOn ANTEI · 18 routes (fal / VolcEngine / BytePlus / Ark)
Strong at
  • Reference endpoint: an image and three voice samples, fed together
  • Every tier from 2.0 to 2.5 is live, with four integrations backing each other up
  • The steadiest first-frame image-to-video here
Weak at
  • Top-end 2.5 is the priciest per second
  • Past 10 seconds you have to stitch
Made here · MiniMax H3
MINIMAXVIDEO · MINIMAX

MiniMax H3

The value workhorse: two regions, and reference input on both.

Dual regionValueReference input
$0.11–0.13 / secondOn ANTEI · 4 routes (overseas + China × reference)
Strong at
  • A reference endpoint even at the low tier
  • The China routes are friendly on latency
  • Well regarded for natural human motion
Weak at
  • Few 1080p tiers
  • Weaker than Kling on stylised camera work
Made here · Wan 3.0
WAN 3.0VIDEO · ALIBABA

Wan 3.0

Alibaba's workhorse: shipped upstream, and we are still adding routes.

Reference inputLow priceGrowing
$0.10 / secondOn ANTEI · 2 routes (text-to-video / reference) · more to come
Strong at
  • The lowest unit price on the whole site, at $0.10 a second
  • The reference-to-video route is live
  • Upstream keeps shipping endpoints and we keep adding them
Weak at
  • Plain camera language
  • Only fair consistency over longer takes
Made here · HappyHorse 1.1
HAPPYHORSEVIDEO · ALIBABA

HappyHorse 1.1

The fast-tier dark horse: top-five value on the blind-test board.

Fast tierBlind-test boardValue
$0.14 / secondOn ANTEI · 1 route
Strong at
  • Top five on the Artificial Analysis blind-test board (v1.1 and v1.0 share the ranking — our note)
  • Quick to finish a clip
  • Friendly unit price
Weak at
  • Only fair style control
  • Little written about it
Made here · Kling 3.0 Pro
KLINGVIDEO · KLING

可灵

The camera one: the name people trust for movement and texture.

Camera workTextureOmni
$0.14–0.17 / secondOn ANTEI · 2 routes (3.0 Pro / O3 Pro Omni)
Strong at
  • First rank for camera movement and physical texture
  • O3 Pro Omni is the all-round endpoint
  • Delicate human performance
Weak at
  • 3.0 Pro does not take reference input (our note)
  • Upper-middle unit price
Gemini Omni Flash sample outputMade here · GPT Image 2
VIDEO · GOOGLE

Gemini Omni Flash

The quick all-rounder: get the idea moving, cheaply.

All-roundFastUsage-based
Usage-basedOn ANTEI · 1 route
Strong at
  • Text-to-video and image-to-video in one place
  • Fast and cheap — right for a rough pass
  • Strong image understanding from the Google side
Weak at
  • A lower ceiling than the dedicated video models
  • Coarse style control
1 / 6 · Seedance
Fish Audio S2.1 Pro sample outputfish.audio official artwork
FISH AUDIOAUDIO · FISH AUDIO

Fish Audio S2.1 Pro

The workhorse for lines: the name people trust for natural Chinese speech.

Chinese dialogueVoice libraryFree / paid options
Free tier $0; paid tier $15 / million UTF-8 bytesOn ANTEI · S2.1 Pro free / paid tiers
Strong at
  • First rank for natural spoken Chinese
  • Pick from the in-app voice library, or clone
  • S2.1 Pro free and paid tiers connected
Weak at
  • Slightly flat English accent
  • Extreme emotion needs the temperature adjusted
ElevenLabs sample outputelevenlabs.io official artwork
ELEVENLABSAUDIO · ELEVENLABS

ElevenLabs

Effects and score: SFX and Music, two routes.

Sound effectsScoreTwo routes
Usage-based – $0.15 / secondOn ANTEI · 2 routes (SFX v2 + Music v2)
Strong at
  • SFX: one sentence, one foley take
  • Music: the structure of a full cue stays under control
  • The industry benchmark for audio quality
Weak at
  • Only fair with Chinese lyrics
  • Billed per second, so length has to be watched
1 / 2 · Fish Audio S2.1 Pro
Rodin Gen-2.5 sample outputhyper3d.ai official artwork
RODIN3D · HYPER3D

Rodin Gen-2.5

Multi-view control: up to 5 references feed a single model.

Multi-viewControllableProduction feel
Usage-basedOn ANTEI · 1 route
Strong at
  • Up to 5 reference images (our cap)
  • Topology clean enough to keep editing afterwards
  • A four-view workflow
Weak at
  • On the slow tier for generation
  • Only fair texture ceiling
Hunyuan3D sample outputMade here · Hunyuan3D render
HUNYUAN3D · TENCENT

Hunyuan3D

The polygon king: a 1.5-million-face ceiling for the detail-minded.

High polyDetailTwo versions
Usage-basedOn ANTEI · 2 routes (v3.1 Pro / v3)
Strong at
  • 40k to 1.5M faces, adjustable (500k by default)
  • Textured mesh in one pass
  • v3.1 Pro sharpens the detail further
Weak at
  • High-poly exports are large files
  • Better at organic forms than hard surfaces
Trellis 2 sample outputmicrosoft.github.io/TRELLIS official teaser
TRELLIS3D · MICROSOFT

Trellis 2

Every dial exposed: resolution, texture and decimation are all yours.

Full controlDecimationGame pipeline
Usage-basedOn ANTEI · 1 route
Strong at
  • Three resolution tiers: 512 / 1024 / 1536
  • Textures from 1K to 4K
  • Decimation from 5k to 2M drops straight into a game pipeline
Weak at
  • Middling speed
  • Stylised input trips it up
TripoSR sample outputtripo3d.ai official artwork
TRIPO3D · TRIPO

TripoSR

The seconds tier: one image, one grey model, a few seconds.

SecondsLightweightMockups
Usage-basedOn ANTEI · 1 route
Strong at
  • The fastest model here
  • Light grey models, right for a mockup
  • Open-source roots, friendly ecosystem
Weak at
  • No texture detail
  • Complex structures blur
1 / 4 · Rodin Gen-2.5

Generating is where it starts.

Generate, archive and reuse in one place. Every result keeps its key settings.

Open the canvas
© 2026 ANTEITerms of service · Privacy policy · Content guidelines
OpeningFeaturesModelsClosing