Best practices for training custom models

Last updated on Sep 21, 2026

Learn how to prepare and review your training dataset to create high-performing custom models in Adobe Firefly.

Custom Models is currently in beta for individual and teams users in Adobe Firefly. Try it now and share your feedback.

Your custom models learn from every image in the training dataset. Selecting consistent, high-quality images and refining captions and tags helps improve the quality and consistency of generated results.

Before training your model, follow these best practices to prepare your dataset.

Select images that showcase your style

Choose images that clearly represent the style or subject you want your custom model to learn.

Model type	Availability	Include	Avoid
Illustrations	Individual and Enterprise	Consistent style and color palette Balanced composition and visual hierarchy Variety in perspective and framing Clear details without visual clutter	Inconsistent styles or mixed techniques Low-quality or incomplete illustrations Limited variety in perspective and framing Overly complex backgrounds that distract from the subject Unrelated or off-theme visuals
Character	Individual and Enterprise	Same character in all images Consistent style and rendering quality Accurate anatomy and consistent proportions Variety of poses, angles, and expressions Clear subject focus without visual clutter	Multiple characters in set Inconsistent styles or mixed techniques Low-quality or incomplete illustrations Limited variety of poses, angles, and expressions Distracting backgrounds Unintentionally repetitive or unrelated elements
Iconography	Individual and Enterprise	High-quality icon illustrations with consistent rendering Consistent style with same stroke weight and corner roundness Clear details without visual clutter Variety of universal symbols	Low-quality or incomplete icon illustrations Inconsistent styles, stroke weights, or corner roundness Distracting backgrounds or unrelated elements Non-universal symbols or limited variety
Photography	Individual and Enterprise	Distinct, consistent visual or lighting style Clear, in-focus, non-human subjects Variety of subjects, angles, and compositions Balanced composition and visual hierarchy Simple or softly blurred backgrounds	Inconsistent photography styles Harsh filters or extreme color grading Blurry or pixelated images Full-body images or shots where human elements like faces, hands, or feet are prominent Overcrowded scenes or distracting backgrounds
Lifestyle photography	Enterprise only	Clear, in-focus people Natural lighting and authentic expressions Variety of poses and compositions Simple or softly blurred backgrounds	Blurry or pixelated images Harsh filters or extreme color grading Overcrowded scenes or distracting backgrounds Full-body or group shots
Photoshoot of a person	Enterprise only	Sharp, well-lit close-ups and mid-distance shots Variety of poses, expressions, and outfits Consistent lighting and environment Clean or softly blurred backgrounds	Faces too small or partially obscured Heavy shadows or harsh lighting Too many similar shots Blurry or low-quality images Full-body or group shots
Isometric and 3D graphics	Enterprise only	Consistent perspective and proportions Cohesive style, lighting, and rendering quality Variety of compositions and angles Clear, uncluttered designs	Low-quality or incomplete renders Inconsistent styles or perspectives Limited variety in angles or compositions Distracting elements or unrelated objects Overly specific colors and design elements, such as exact proportions
Explore brand expression	Enterprise only	Strong, consistent brand style throughout Clear compositions with room to breathe Expressive, on-brand characters and scenes Clean rendering with balanced lighting	Mixed styles or inconsistent perspectives Crowded scenes with unclear focus Off-brand props or unrelated visuals Incomplete or low-quality illustrations
Backgrounds for product shots	Enterprise only	Visually distinct and well-executed concepts Consistent structure, lighting, and detail Strong form and clear silhouette High-quality images with clean rendering	Repeated shapes or minor variations Distracting backgrounds or details Incomplete or low-quality renders Mixed rendering styles or effects

Use high-quality training images

Before uploading images, verify that every image meets the following requirements.

Requirement	Guidance
Format	JPG or PNG
File size	Less than 20 MB for individual and team users and Less than 50 MB for enterprise users.
Resolution	1024 × 1024 pixels or larger
Aspect ratio	Keep a consistent aspect ratio across the dataset. Maximum 16:9 for landscape or 9:16 for portrait.
Quantity	Upload 10–30 images.
Variety	Include different perspectives, backgrounds, and compositions while maintaining a consistent visual style.
Subject	Focus on the primary style or subject. Crop out unnecessary elements where possible.
Background	Avoid transparent backgrounds and distracting elements.
Consistency	Avoid unintended repeating props, accessories, or backgrounds unless they are intended to become part of the style.

Note

If you’re using Firefly Custom Models as an enterprise user and training models for organizational use, ensure the dataset reflects approved brand assets and follows your organization’s governance policies.

Review and refine captions

Firefly automatically generates captions for every uploaded image. Review each caption before training.

Do	Don’t
Correct inaccurate captions.	Leave incorrect auto-generated captions unchanged.
Add visual details that describe your style or subject.	Use vague descriptions.
Use varied sentence structures across captions.	Repeat identical caption patterns throughout the dataset.
Describe recognizable people or landmarks with context.	Assume Firefly understands proper names without context.
For subject models, include the Concept ID in every caption.	Omit the Concept ID from subject-based datasets.

Review model tags

Model tags describe characteristics that should remain consistent across all generated images.

Include	Avoid
Permanent characteristics such as watercolor texture, flat illustration, brown hair, soft lighting, or geometric style.	Temporary attributes such as expressions, clothing, props, accessories, or objects being held.

Complete the training checklist

Before selecting Train, verify the following.

Training images	Captions	Model tags
10–30 images uploaded	Captions reviewed for accuracy	At least three tags added
JPG or PNG format	Specific, descriptive language	Tags describe permanent characteristics
Images meet size and resolution requirements	Concept ID included for subject models	Temporary attributes excluded
Consistent visual style	Varied sentence structure
Variety in composition and perspective	Recognizable subjects described clearly
No distracting backgrounds or repeating elements

Turn text into art with Adobe Firefly
Create beautiful images and video clips from text prompts using generative AI.

Open the app

Create and train custom models

Generate and edit images with custom models

Learn

Learn with step-by-step video tutorials and hands-on guidance right in the app.

Community

Join discussions, find answers, learn from experts, and share your knowledge.

Adobe Home

Access your favorite Creative Cloud apps, services, file management, and more.