Skip to main content
Reference-to-Image models generate a new image from one or more reference images combined with a text prompt. In Automat Studio, these models power the “from reference” workflow used when creating assets — building a character, location, prop, or wardrobe item that matches the look of images you provide.
These models are used in Character Studio, Location Studio, Prop Studio, and Wardrobe Studio when you generate an asset from reference images.
Your reference is not copied. Automat generates a new image from it — a studio portrait for a character, a cinematic wide shot for a location, a product-style shot for a prop, and a ghost-mannequin shot with nobody wearing it for a wardrobe item. Expect a close likeness, not an exact reproduction of the file you uploaded. That is what makes the asset reusable: it has to be re-posed, re-lit and placed into shots that don’t exist in your original photo.

​
When to Use

  • Asset from References — Create a character, location, prop, or wardrobe item that matches uploaded reference images
  • Consistency — Keep a recurring asset visually consistent across your project
  • Likeness Matching — Generate variations that preserve the identity or style of your references
  • Concept Realization — Turn moodboard or inspiration images into a usable asset
These models consistently deliver the best results for reference-based generation:

​
Google Nano Banana 2

  • Credits: 11
  • Rating: ⭐⭐⭐⭐⭐ (5/5)
  • Provider: Google
  • Avg. Duration: ~18 seconds
  • Best for: Premium quality with strong reference adherence (Default model)

​
OpenAI GPT Image 2

  • Credits: 27
  • Rating: ⭐⭐⭐⭐⭐ (5/5)
  • Provider: OpenAI
  • Best for: Excellent prompt comprehension for demanding reference matching

​
OpenAI GPT Image 2.5 Flare

  • Credits: 9
  • Rating: ⭐⭐⭐⭐⭐ (5/5)
  • Provider: OpenAI
  • Best for: Strong likeness preservation from a reference photo, quickly

​
OpenAI GPT Image 2.5 Sunburst

  • Credits: 9
  • Rating: ⭐⭐⭐⭐⭐ (5/5)
  • Provider: OpenAI
  • Best for: Tighter control when the reference has a complex subject or background; slower than Flare

​
Google Nano Banana

  • Credits: 6
  • Rating: ⭐⭐⭐⭐ (4/5)
  • Provider: Google
  • Avg. Duration: ~12 seconds
  • Best for: Fast, high-quality generation at low cost

​
Kling 3.0 Image

  • Credits: 5
  • Rating: ⭐⭐⭐⭐ (4/5)
  • Provider: fal.ai
  • Avg. Duration: ~38 seconds
  • Best for: Affordable, high-quality reference-based generation

​
Upload Requirements

One reference image per asset, in JPG, PNG, or WebP, up to 10MB.

​
Tips for Best Results

  1. Use Clear References — Provide high-quality, well-lit reference images that clearly show the subject
  2. Combine with Prompts — A descriptive prompt sharpens what the model takes from your references
  3. Match the Goal — Use references that reflect the angle, framing, and style you want in the result
  4. Iterate — Generate multiple versions and keep the ones that best match your references
  5. Keep Assets Consistent — Reuse the same references when generating variations of a recurring asset
  6. Push Likeness with Final — For a real face, a sharp, well-lit, front-on photo on Final quality holds the likeness noticeably better than Draft. Lock the reference image you are happy with before you shoot any shots

Character Studio

Create characters from reference images

Wardrobe Studio

Build wardrobe items from references