Feed AI - Choosing the Right Model

Learn about the text and image AI models available in Feed AI, their technical capabilities, credit consumption, and how to pick the right model for your catalog needs.

Table of Contents

1. Overview
2. Text Models: Which One to Choose
3. Image Models: Which One to Choose
4. Key Features & Capabilities
5. Quick Decision Guide
6. Related Articles

Overview

When creating a task in Feed AI, you choose between Text and Image processing. This selection is fixed at creation time and determines which models are available:

• Text tasks: Rewrite titles, optimize descriptions, extract attributes, or categorize products.
• Image tasks: Generate lifestyle staging, create product variations, or produce clean packshots.

All models are available by default across all contracts.

Note on credit estimation: Before launching any task, SmartFeeds calculates an upfront credit estimate based on your catalog size, selected model, and input data. You are only charged for products successfully processed. If an item fails, no credits are deducted.

Text Models: Which One to Choose

* Estimates assume an average of 1,000 input tokens and 400 output tokens per product. Figures in parentheses include reading a 1024×1024 product image.

Recommendations:

• Default choice: Use GPT-6 Luna. It offers the best performance-to-credit ratio and the highest throughput for large product feeds.
• European Data Sovereignty:
• • For sovereign EU hosting at the lowest credit consumption, pick Mistral Small 3.1 (hosted on Vertex AI in Europe, data never leaves the EU). Note: Due to lower rate limits (300 req/min), reserve it for small catalogs and straightforward text generation.
  • For large catalogs requiring EU hosting with high throughput, choose Gemini 3.5 Flash Lite.

Image Models: Which One to Choose

Recommendations:

• Default choice: Use Flare set to medium quality. It delivers fast rendering, optimized credit consumption (~88 credits for 1,000 images), and native background cutout support.
• Premium visuals: Use Sunburst when maximum image fidelity is required (e.g., hero banners or small premium collections).
• High Resolution: Use Nano Banana 2 when native 2K/4K resolution is mandatory.

Key Features & Capabilities

Feed AI incorporates built-in technical safeguards and optimizations:

1. Multimodal Vision & Input Fidelity

All text models (including GPT-6 Luna, Gemini, and Mistral Small 3.1) and GPT-based image models can directly inspect source product images:

• Accurate details: The model looks at your product photos (colors, materials, shapes) to write accurate descriptions without guessing.
• Credit cost: Scanning a product photo adds a small cost (~0.8 extra credit per 1,000 products on GPT-6 Luna)

2. Guaranteed Structured Output

• Text Tasks: Formatting strictly matches your configured schema. Every requested attribute returns in its dedicated column with 100% reliability.
• Dual Output on Image Tasks (Flare & Sunburst): A single task execution can generate the new product image and extract accompanying text fields (such as image alt-text, color tags, or SEO labels) simultaneously.
• Note: Nano Banana 2 creates images only, without accompanying text fields.

3. Native Transparent Backgrounds

Flare and Sunburst can automatically remove the background (transparent PNG). You get clean product cutouts without needing Photoshop or extra tools.

4. Aspect Ratios & Variations

• Supported formats: Square (1:1 / 1024×1024), Landscape (1536×1024), and Portrait (1024×1536).
• Generate between 1 and 3 visual variations per product in a single run.

Quick Decision Guide

• Standard catalog rewriting at scale? → Choose GPT-6 Luna.
• Strict EU data sovereignty on a budget? → Choose Mistral Small 3.1.
• High-volume EU data residency? → Choose Gemini 3.5 Flash Lite.
• E-commerce lifestyle staging or cutouts? → Choose GPT Image 2.5 – Flare (medium).
• Top-tier hero campaign visuals? → Choose GPT Image 2.5 – Sunburst (high or max).

Related Articles

For further assistance with SmartFeeds, consider reviewing these articles:

• Join
• Mapping
• Feed AI : Getting Startted

For additional help, feel free to reach out via our Contact Us page.