Intelligence

The intelligence layer eliminates guesswork in model selection, parameter configuration, and workflow planning.

Model Recommendations

The recommend tool uses a preference scoring system to surface the best model for any task. Scores reflect community consensus and real-world performance: a +100 score marks the default choice, a +50 marks a strong alternative, and a −30 deprioritizes models that have been superseded. The table below shows the default model for each category and a brief rationale.

CategoryDefault ModelRationale
image generationmodel_google-gemini-3-1-flashGemini 3.
image editingmodel_google-gemini-3-1-flashGemini 3.
fast imagemodel_p-imageP-Image is the fastest image model.
text renderingmodel_google-gemini-3-1-flashGemini 3.
pixel artmodel_retrodiffusion-plusRetrodDiffusion Plus is the specialist for high-quality pixel art at native 256x256.
vector illustrationmodel_recraft-v4-svgRecraft V4 SVG generates native vector output.
vectorizationmodel_recraft-vectorizeConverts raster images to SVG.
background removalmodel_photoroom-background-removalPhotoroom is the best default for BG removal.
upscalingmodel_topaz-image-upscaleTopaz is best for photos — face enhancement, natural detail recovery, noise handling.
style transfermodel_google-gemini-3-1-flashEdit with Prompts using Gemini 3.
reframe imagemodel_scenario-gemini-reframeGemini Reframe for images.
reframe videomodel_luma-reframe-videoLuma Reframe for videos.
video t2vmodel_kling-v3-t2v-proKling V3 T2V Pro is the quality default — far superior realism.
video i2vmodel_kling-v3-i2v-proKling V3 I2V Pro is the quality default — far superior realism, supports elements, first+last frame, 3-15s.
video motion controlmodel_kling-v3-pro-motion-controlKling V3 Motion Control Pro transfers motion from a reference video to a character image.
video extendmodel_veo3-1-extend-videoVeo 3.
video upscalemodel_topaz-video-upscaleTopaz for video upscaling to 4K.
video reframemodel_luma-reframe-videoLuma Reframe expands videos in any direction to 6 preset aspect ratios.
lipsyncmodel_veed-fabric-1-0Veed Fabric is excellent for lipsync — reliable, high quality.
image to 3dmodel_hunyuan-3d-pro-3-1-i23dHunyuan 3D 3.
3d partcraftmodel_hunyuan-3d-partBetter workflow: create the 3D model first (Hunyuan 3D 3.
3d retexturemodel_meshy-retextureChange surface appearance while preserving geometry.
3d remeshmodel_hunyuan-polygen-1-5Hunyuan PolyGen 1.
ttsmodel_xai-grok-ttsGrok TTS — hyper fast, super cheap, versatile.
sfxmodel_elevenlabs-sound-effects-v2ElevenLabs SFX v2 for environmental sounds, mechanical effects, organic textures.
music instrumentalmodel_beatoven-music-generationBeatoven for instrumental tracks up to 2m30s at 44.
music with lyricsmodel_minimax-music-2-0MiniMax Music 2.

Platform-Aware Formatting

The intelligence layer automatically detects platform intent from your prompt — mention "TikTok", "Instagram story", or "YouTube thumbnail" and the correct aspect ratio and resolution are applied without manual configuration. The following platforms are recognized:

PlatformAspect RatioResolutionNote
instagram_post4:51080pxInstagram feed post
instagram_story9:161080pxInstagram/TikTok story
instagram_reel9:161080pxInstagram Reel
tiktok9:161080pxTikTok vertical video
youtube_thumbnail16:91280pxYouTube thumbnail
youtube_video16:91920pxYouTube video frame
twitter_post16:91200pxTwitter/X post image
linkedin_post1:11080pxLinkedIn square post
facebook_post1:11080pxFacebook post
facebook_cover16:91640pxFacebook cover photo
pinterest2:31000pxPinterest pin
app_icon1:11024pxMobile app icon
game_asset1:11024pxGame asset (power of 2)
game_texture1:11024pxSeamless game texture
print_a43:42480pxA4 print at 300 DPI
print_poster2:33000pxPoster print
wallpaper_desktop16:92560pxDesktop wallpaper
wallpaper_phone9:161440pxPhone wallpaper
banner_web16:91920pxWebsite hero banner
email_header16:9600pxEmail header image
avatar1:1512pxProfile picture / avatar
ultrawide21:92560pxUltrawide monitor

Character Consistency

For character consistency across scenes, the recommended approach is multi-reference editing — pass the character image as a reference to each scene generation. Gemini 3.1 supports up to 14 reference images, Seedream 4.5 up to 10, GPT Image 1.5 up to 5.

Best models for character consistency

  • model_google-gemini-3-1-flash (14 refs)
  • model_bytedance-seedream-4-5 (10 refs)
  • model_gpt-image-1-5 (5 refs)

LoRA training is an advanced alternative suited for art style replication at scale — 100+ images in a consistent style or brand-specific visual language. It requires 10–20 curated training images and 30+ minutes of training time. For most character consistency needs, the multi-reference editing approach above is faster and easier.

Pipeline Planning

The recommend tool matches your intent against pre-built pipeline templates. When a multi-step workflow is detected — product video, talking head, game asset, character sheet — the tool surfaces the right sequence of tools and models automatically.

product_video— Product image → hero video → voiceover → combine
product videoproduct revealproduct launchcommercialad video
  1. generate — Generate hero product image
  2. edit — Clean transparent product image
  3. generate — Image-to-video product reveal
  4. generate — Generate voiceover(optional)
  5. generate — Generate background music(optional)
talking_head— Portrait → voice → lipsync video
talking headspokespersonpresenteravatar videowelcome videoapp intro
  1. generate — Generate portrait/headshot
  2. generate — Generate voiceover
  3. generate — Lipsync portrait + audio → talking video
game_asset_3d— Concept art → clean background → upscale → 3D conversion
game asset 3d3d game asset3d model for gamegame prop 3d
  1. generate — Generate concept art (white bg, isometric)
  2. edit — Remove background for clean 3D input
  3. edit — Upscale 2x for better 3D texture quality(optional)
  4. generate — Convert to 3D model
character_sheet— Reference character → multiple scenes with reference editing
character sheetcharacter turnaroundcharacter referencecharacter consistencysame character
  1. generate — Generate strong reference character image
  2. generate — Scene 1: Edit with reference image + scene prompt
  3. generate — Scene 2: Edit with same reference + new scene
  4. generate — Additional scenes as needed
logo_to_vector— Generate logo → vectorize to SVG
logo vectorvectorize logosvg logologo for printscalable logo
  1. generate — Generate logo design
video_chain_long— Chain multiple video clips via lastFrame for longer content
long video30 second video60 second videoextended videomulti-clip
  1. generate — Generate hero image for first frame
  2. generate — Image→video clip 1 (10s)
  3. manage_assets — Get lastFrame asset ID from clip 1
  4. generate — LastFrame→video clip 2 (10s)
  5. manage_assets — Get lastFrame from clip 2, chain more clips as needed
style_transfer— Generate or use existing image → edit with style prompt
style transfermake it look likeghibli styleanime styleoil painting stylewatercolor stylepixel art style
  1. generate — Edit source image with style description