Skip to content
Provider Models

Provider Models

The AI models Milk & Ink Studio orchestrates on your behalf. Each row tells you the model's training posture for the inputs we send it, where it's routed, and the provider's stance on output rights. This page is referenced by our Terms of Service (Section 6) and our AI Rights & Content Policy.

No training

We use the provider's no-training endpoint or the provider does not train on customer inputs by default. Safest posture for IP-sensitive work.

Provider default

Inputs may be retained by the provider per their default policy. Acceptable for general use; review the linked terms before relying commercially.

Opt-in only

Training only happens when explicitly opted in (e.g., training a custom LoRA on your own reference images at your instruction).

Provider ModelCategoryRouted viaTraining postureOutput rightsProvider terms
Wan 2.1
Text-to-video (1.3B draft, 13B final)
Videofal.ai No trainingCustomer owns output (commercial use permitted via fal.ai terms)Link
Seedance 2.0
Image-to-video with native audio
Videofal.ai (ByteDance model) No trainingCustomer owns output (per fal.ai + ByteDance commercial license)Link
Kling 2.5 Turbo / 2.6 Pro / 3.0
Image-to-video, start+end frame interpolation
Videofal.ai (Kuaishou model) No trainingCustomer owns output (per fal.ai + Kuaishou commercial license)Link
Google Veo 3
Cinematic photoreal video with native audio (live-action lane)
Videofal.ai (Google Veo) No trainingCustomer owns output (per Google Veo commercial license, surfaced via fal.ai)Link
Runway Gen-4
Cinematic 1080p video
VideoRunway (direct API) Provider defaultCustomer owns output (per Runway commercial terms)Link
Nano Banana 2
Character-identity locking (image-side)
Imagefal.ai No trainingCustomer owns outputLink
Flux LoRA Training
Custom character LoRAs trained on customer's reference images
Imagefal.ai (Black Forest Labs) No trainingCustomer owns the trained LoRA artifact — workspace-scoped, never used outside the customer's workspaceLink
ElevenLabs (TTS)
Eleven v3 — emotive voice synthesis with audio-tag interpretation
AudioElevenLabs (direct API) No trainingCustomer owns generated audio (per ElevenLabs commercial license)Link
ElevenLabs (Music + Video-to-music)
Music compose, composition_plan, video-to-music re-score
AudioElevenLabs (direct API) No trainingCustomer owns generated audioLink
ElevenLabs (SFX) + fal MMAudio
Prompt-only sound effects (ElevenLabs); video-aware Foley (MMAudio)
AudioElevenLabs / fal.ai No trainingCustomer owns generated audioLink
fal Stable Audio
Stable Audio Open — 47s music clips (fallback when ElevenLabs unconfigured)
Audiofal.ai No trainingCustomer owns generated audio (Stable Audio Open license)Link
Sync.so
Lipsync — aligns generated audio to generated video
AudioSync.so (direct API) No trainingCustomer owns the lipsynced outputLink
Groq
llama-3.3-70b — script planner, suggestions, copilot (default)
LLMGroq (direct API) No trainingCustomer owns generated textLink
Google Gemini (AI Studio)
Vision + LLM — continuity scoring, multimodal review (default)
LLMGoogle AI Studio (direct API) No trainingCustomer owns generated assessmentsLink

What we don't promise about Provider Models

  • No copyright guarantee. The U.S. Copyright Office and many other jurisdictions limit copyright protection for content generated solely by AI without sufficient human authorship. We make no representation about whether any Output is copyrightable. For commercial registration, document the human creative choices that went into the work (canon decisions, prompts, edits, selections, post-production).
  • Providers can change. We may add, remove, or substitute Provider Models without notice if a provider deprecates a model or changes its terms. We surface the current set in this table and in the provider picker inside the Studio composer.
  • Enterprise BYOK. Enterprise customers can route generation through their own provider keys via signed delegated-key receipts. Talk to enterprise@milkink.studio.