18Models
DocsModelPricingUse CasesLog inTry for Free
Try for FreeLog in

Character consistency without a LoRA

Reference-to-video · multi-subject references · storyboard mode · 12-panel character sheets · no training step

Consistent characters are the difference between a demo and a product — and most teams reach for LoRA training to get them. 18models does it with references instead: Wan 2.7 reference-to-video accepts up to five subject references (images or clips, each with an optional voice sample), HappyHorse r2v up to nine reference images, Wan 3.0 up to ten images plus video and audio, and Wan 2.7 image produces 12-panel character sheets to seed them all. No training, no per-character upkeep, NSFW allowed under the content policy.

The toolkit

NeedModel / modeReferences acceptedNotes
Same character in a video clipwan2.7-r2v-multiple-uncensored1–5 images and/or videos, each may include voice720p $0.113/s · 1080p $0.169/s; 2–15 s (10 s with video refs); optional first frame via SET_FIRST_IMG_AS_FIRST_FRAME
Same character, cheaper / verticalhappyhorse-1.1-r2v-uncensored1–9 images480p $0.048/s · 720p $0.097/s · 1080p $0.125/s; 3–15 s; 9 aspect ratios
Character + voice + motion reference, longerwan3.0-video-reference-uncensoredup to 10 images, 5 videos, 5 audio (≤15 s each in total)from $0.058/s; 2–30 s; refer to "Image 1 / Video 1 / Audio 1" in the prompt
Whole scene laid out panel by panelwan2.7-r2v-storyboard-uncensored1 multi-panel imageRenders the storyboard as one continuous clip
Character sheet to seed everythingwan2.7-image-grid-t2i / grid-i2itext, or 1 imageUp to 12 consistent panels; $0.034/panel Standard
New still of the same characterqwen-image3-edit-uncensored / wan2.7-image-edit-uncensored1–3 / 1–9 imagesOutfit, pose, scene changes with the face intact
Animate an approved still*-i2v-firstframe (Wan 2.7, HappyHorse, Wan 3.0, Seedance)1 imageThe still becomes frame one

A workflow that holds up

  1. Canonical portrait. One Pro-quality image (Qwen-Image 3.0 Pro or Wan 2.7 image Pro): clean lighting, neutral background, three-quarter view.
  2. Character sheet. wan2.7-image-grid-i2i-uncensored from that portrait → 12 panels (front / side / expressions / outfits). Keep 3–5 panels that best show face, body and clothing.
  3. Stills. For every new image, edit from the canon (Qwen-Image edit with 1–3 of those panels) instead of prompting from scratch.
  4. Video. Send the same panels as images to Wan 2.7 r2v (add a voice sample if the character speaks) or HappyHorse r2v; for scripted multi-shot scenes, lay out a storyboard grid and use storyboard-to-video.
  5. Continuity. Extend a clip with wan2.7-i2v-continuation-uncensored (source clip + optional target last frame) instead of regenerating.
{
  "model": "wan2.7-r2v-multiple-uncensored",
  "prompt": "The woman from image 1 sits on the bed and speaks the line in a low voice, soft window light",
  "images": [
    { "url": "https://cdn.example.com/char-front.jpg", "voice": "https://cdn.example.com/char-voice.mp3" },
    "https://cdn.example.com/char-side.jpg"
  ],
  "params": { "resolution": "720p", "ratio": "9:16", "duration": 6 }
}

Limits, honestly

  • No LoRA / checkpoint hosting today; reference-based consistency covers most companion, game and creator needs and needs no training step. LoRA support is being evaluated; nothing to announce yet.
  • References of real, identifiable people are prohibited for NSFW generation and screened on input — the canon must be a fictional character.
  • Video references cap Wan 2.7 r2v at 10 seconds; total video/audio reference length is limited on Wan 3.0 (15 s each) and Seedance (30 s each).

Related

  • Wan 2.7 video API · HappyHorse · Wan 3.0 · Wan 2.7 image
  • API for AI companion apps · API for adult games

Prices are USD base list prices from /pricing; the effective price for your key is returned by GET /ent/v2/prices. NSFW output is permitted under the Content Policy — fictional adults only, minors and real people prohibited and screened.

FAQ

How do I keep a character consistent in AI video without a LoRA?

Use reference-to-video: Wan 2.7 r2v takes 1–5 reference images/videos (each may carry a voice sample) and keeps the subject across the clip; HappyHorse r2v takes 1–9 reference images; Wan 3.0 reference mode mixes up to 10 images, 5 videos and 5 audio clips.

Does 18models support LoRA or custom checkpoints?

Not today. Reference-based consistency covers most companion, game and creator use cases; LoRA hosting is being evaluated; any change will be announced via the changelog.

What is storyboard-to-video?

Wan 2.7 storyboard mode takes one multi-panel image (for example a 12-panel grid from Wan 2.7 image) and renders it as one continuous clip, keeping the character and scene consistent panel to panel.

How do I get a good reference image?

Generate a clean, well-lit portrait or three-quarter shot (Qwen-Image 3.0 Pro or Wan 2.7 image Pro), then produce a character sheet with Wan 2.7 grid-from-image and pick the panels that best show face, body and outfit as references.

Build on a provider that says yes in writing. One API key, 34 models, USD prices, and a content policy you can quote.
Get your API keyRead the docs
18Models

High-fidelity AI video generation built for platforms and creators who push boundaries.

All systems operational
Product
DocsPricingAll modelsPlaygroundNSFW policy trackerAPI StatusChangelog
Models
Wan 2.7 videoWan 3.0 videoHappyHorseSeedance 2.5Qwen-Image 3.0Z-Image TurboWan 2.7 image
Solutions
AI companion appsAdult gamesCreator platformsCharacter consistencyMigrate from Novita
Compare
vs Novitavs WaveSpeedvs fal.aivs Replicatevs CivitAIvs ModelsLabvs Segmind
Legal
Privacy PolicyTerms of ServiceContent PolicyCookie Policy
© 2026 18models.ai. All rights reserved.
TLS in transit · Hashed API keys · Operational audit logs