AI Model Selection Guide

Last Updated: 2026-08-17

Overview

This guide helps you choose the right image model, video model, or processing tool for each task on WeShop AI.

Supported image models: GPT Image 2, Nano 2, Nano Pro, Seedream 5 Pro, Seedream 5.0 Lite, Midjourney, and Z-Image. Supported video models: Kling V3, Kling V3 Omni, MiniMax H3, Seedance 2.5, Seedance 2.0, and Seedance Mini.

For API fields and Agent parameters, see API Reference and Agent. If you have any questions, please email: hi@weshop.ai.

Usage Principles

Assess the Task Before Choosing a Model

Model selection should follow this order:

1. Determine whether the task is pure text-to-image, reference-image generation, local editing, or multi-image consistency. Midjourney and Z-Image support text-to-image only; exclude both whenever image input or image-editing is required.
2. Define the delivery tier: draft, internal review, live campaign, or premium commercial delivery.
3. Confirm core constraints: subject consistency, Asian aesthetics, photographic quality, speed, cost, or batch scale.
4. Choose the lowest-cost model that satisfies the core constraints.
5. On generation failure, optimize inputs first, then upgrade the model; do not rely on repeated blind retries.

Default Recommendations

Relative Price Order

Under the same pricing basis, this guide uses the following relative price order (high to low):

GPT Image 2 High
> Nano Pro
> Nano 2 = Seedream 5 Pro
> GPT Image 2 Medium = Seedream 5.0 Lite
> GPT Image 2 Low

The tiers above reflect relative invocation cost only; they are not equivalent to model quality ranking. Actual total cost is also affected by output dimensions, generation count, retry count, and platform billing rules.

Midjourney and Z-Image share the same per-image price. Their relative pricing against other P1–P5 models is not yet confirmed, so they are not included in the tiers above. Midjourney always outputs 4 images per call; when calculating per-invocation cost, bill for 4 images, not a single image.

Capability Declaration Rules

• This guide describes recommended usage boundaries; it does not represent official hard limits of the underlying models.
• GPT Image 2 supports Low, Medium, and High tiers. Other models and tools do not support the quality parameter.

Hard Routing Rules for Text-in-Image Tasks

• Whenever the output image must include titles, poster copy, packaging text, signage, menus, UI text, or other readable characters—including Chinese—route uniformly to GPT Image 2 Medium.
• Multilingual generation (including Chinese-inclusive combinations), image text translation, language replacement, and localized layout also route fixedly to GPT Image 2 Medium.
• Z-Image is not a candidate for Chinese text requirements; consider Z-Image only for pure text-to-image tasks that explicitly require Chinese cultural elements.
• Nano 2, Nano Pro, Seedream 5.0 Lite, and Seedream 5 Pro are not primary or fallback models for text-in-image generation.
• GPT Image 2 tiers follow a progressive rule: Low for text layout validation; Medium as the default for routine text and translation deliverables; High only when Medium fails acceptance or the user explicitly requests it.
• Text content must be given verbatim in the prompt, with explicit language, casing, placement, font style, and typographic hierarchy.
• If text accuracy still does not meet delivery requirements, switch to “text-free base image + layout tool for text”; do not blindly switch among models that lack text input capability.

Hard Limits for Text-to-Image-Only Models

• Midjourney and Z-Image perform text-to-image only; input must consist solely of text prompts and generation parameters.
• When any reference image, product image, portrait, background, mask, first/last frame, or source image for editing is present, do not call Midjourney or Z-Image.
• Model swap, outfit change, virtual try-on, multi-product compositing, subject consistency, inpainting, expand image, background removal, and image translation are not pure text-to-image tasks; handle them with the corresponding image models or tools.
• Midjourney and Z-Image do not support quality; do not pass this field in requests.
• Midjourney always outputs 4 images per call; the invocation layer must not request a different count, and the business layer must receive, bill, and filter all 4 results.
• Unit pricing, relative pricing against other models, and speed for both models are not yet confirmed; Z-Image’s unified quality tier is also pending confirmation. Complete configuration before integration; do not infer values.

Seedream Series Usage Limits

• Consider Seedream 5.0 Lite or Seedream 5 Pro only when the task has high lighting requirements or explicitly requires Asian aesthetics.
• For general generation, subject consistency, routine product shots, and standard commercial images that do not meet either condition above, default to GPT Image 2 or the Nano series.
• Seedream 5.0 Lite is for routine deliverables; Seedream 5 Pro is for the same class of needs with higher lighting, material, or final-delivery requirements.
• Text-in-image rules take precedence: whenever readable text, multilingual output, or translation is required—including Chinese—use GPT Image 2 Medium.

GPT Image 2 Tier Hard Rules

GPT Image 2 must use progressive tier selection. Do not jump to High simply because the task description mentions “multiple images,” “subject consistency,” “commercial image,” or “final delivery.”

Enforced sequence:

1. Call Medium first for routine tasks; for batch tasks, generate 1 test image with Medium first.
2. After the test image passes identity, outfit, composition, text, or product-structure acceptance, complete remaining outputs with Medium.
3. Upgrade to High only when Medium results are actually substandard or the user explicitly requests High.
4. High cannot compensate for missing reference assets. If outfit, subject angle, or product structure is not visible in references, add assets first.
5. When upgrading to High, record the specific reason—for example, “Medium subject detail failed acceptance”—not merely “high quality required.”

Image Models

GPT Image 2

Basic Information

Recommended Use Cases

Tier and Alternative Boundaries

For multilingual tasks, call and accept each target language separately; do not output multiple language versions in one request.

Nano 2

Basic Information

Recommended Use Cases

• Batch generation of visual directions and composition candidates.
• Early A/B testing for ad creative.
• Social content sketches, storyboards, and mood boards.
• Tasks needing fast batch exploration with quality above the lowest-cost draft tier.
• Filter prompts, then redo on a higher-quality model.

Not Recommended as Primary

Nano Pro

Basic Information

Recommended Use Cases

Note

Readable text in images: do not call Nano Pro; route uniformly to GPT Image 2 Medium.

Boundary with Nano 2

Seedream 5 Pro

Basic Information

Recommended Use Cases

• Commercial photography with high requirements for light direction, tonal hierarchy, shadow transitions, rim light, or ambient light.
• Automotive, jewelry, watch, fragrance, and luxury ads requiring studio lighting and texture sculpting.
• Product shots where glass, metal, leather, and similar materials need fine lighting expression.
• Asian subjects, Asian fashion, beauty, and localized e-commerce visuals.
• Brand key visuals emphasizing both Asian aesthetics and professional studio lighting.

Not Recommended as Primary

Seedream 5.0 Lite

Basic Information

Recommended Use Cases

Note

Readable text in images: do not call Seedream 5.0 Lite; route uniformly to GPT Image 2 Medium.

Note

If the task does not have high lighting requirements and does not emphasize Asian aesthetics, do not call Seedream 5.0 Lite; use GPT Image 2 or the Nano series instead.

Boundary with Seedream 5 Pro

Midjourney

Basic Information

Recommended Use Cases

Prohibited Scenarios

Midjourney always generates 4 images. Even when the business needs only 1, receive all results and filter afterward.

Z-Image

Basic Information

Recommended Use Cases

Prohibited Scenarios

Video Generation Models

Video model pricing, quality tiers, supported duration, resolution, and concurrency limits are not yet confirmed. The following defines model boundaries and invocation structure only; concrete enums must be overridden by actual platform configuration.

Kling V3 / V3 Omni

Basic Information

Kling V3 Recommended Use Cases

Kling V3 Omni Recommended Use Cases

Version Boundaries

Not Recommended as Primary

MiniMax H3

Basic Information

Recommended Use Cases

Not Recommended as Primary

Required visual input combinations, duration, and format for audio-driven generation must follow the actual platform; do not assume pure-audio requests always work.

Seedance 2.5 / 2.0 / Mini

Basic Information

Version Selection

Recommended Use Cases

• Music videos, mood shorts, and artistic social content.
• Non-photorealistic, emotional, or stylized visual expression.
• Shorts centered on rhythm, color, transitions, and visual mood.
• Seedance Mini for validating style, rhythm, and camera direction.

Not Recommended as Primary

Video Model Quick Decision

Does the task require large-amplitude motion or high-dynamic action?
├─ Yes → MiniMax H3
└─ No: Does it require multi-image, video, or complex references?
   ├─ Yes → Kling V3 Omni
   └─ No: Does it require precise first/last frame or controllable product showcase?
      ├─ Yes → Kling V3
      └─ No: Is audio-visual sync or artistic expression the core goal?
         ├─ Yes → Seedance 2.5
         └─ No: Is it only lightweight concept validation?
            ├─ Yes → Seedance Mini
            └─ No → Seedance 2.0

Video Routing Priority

1. Precise first/last frame and product structure control take priority over artistic style → Kling V3.
2. Multi-reference input takes priority over standard first/last frame → Kling V3 Omni.
3. Large-amplitude motion takes priority over identity and first/last frame control → MiniMax H3.
4. When audio-visual sync and artistic expression are prioritized → Seedance 2.5.
5. Seedance 2.0 for routine artistic shorts; Mini for lightweight validation.
6. When conflicting goals coexist, split shots or have the user confirm the highest priority first.

Image Processing Tools

Use processing tools when the task is not image or video generation. For API parameters, see Agent.

Tool Overview

Virtual Try-On Routing Rules

Do not call Virtual Try-On when:

When Virtual Try-On does not apply, route to image models per this guide: GPT Image 2 for text or multilingual tasks; Seedream series when lighting requirements are high or Asian aesthetics apply; otherwise GPT Image 2 or Nano series by consistency, quality, speed, and price.

Tool Selection Rules

Need to remove the original background?
├─ Yes → Remove BG
└─ No: Is it virtual try-on, model swap, or background swap?
   ├─ Yes: Is there exactly 1 product image, at most 1 single-person model image, and at most 1 background image?
   │  ├─ Yes → Virtual Try-On
   │  └─ No → Image model workflow
   └─ No: Need to change aspect ratio, platform size, or crop/composition?
      ├─ Yes → Expand Image
      └─ No → Image generation or editing workflow

Model Selection & Routing

Image Model Comparison Matrix

“Output quality positioning” in this table uses internal assessment criteria; it is not a request parameter and does not equal fit for a specific style. Only GPT Image 2 supports the quality parameter. Seedream enters the candidate set only for tasks with high lighting requirements or Asian aesthetics. Midjourney and Z-Image support text-to-image only; they share the same per-image price, and Midjourney always outputs 4 images per call.

Image Model Quick Decision Tree

Does the task include any image input, reference image, or image-editing requirement?
├─ Yes → Exclude Midjourney and Z-Image
│  ├─ Image translation, multilingual, or readable text editing → GPT Image 2 Medium
│  └─ Other tasks → Choose GPT Image 2, Nano, or Seedream series by consistency, cost, lighting, and Asian aesthetics
└─ No (pure text-to-image)
   ├─ Artistic creation, inspiration, or illustration → Midjourney
   ├─ Any readable text or multilingual (including Chinese) → GPT Image 2 Medium
   ├─ Chinese cultural elements → Z-Image
   ├─ Photorealism as core without artistic emphasis → Z-Image
   └─ Other general tasks → GPT Image 2 Medium; then adjust by price, batch, and style needs

Recommended Image Production Pipelines

Model Fallback

When the primary model does not meet requirements, use the following fallback chains.

Image Model Fallback Chain

Readable text, multilingual, translation, and image text editing tasks do not use the cross-model fallback chain above; default to GPT Image 2 Medium.

Video Model Fallback Chain

Do not silently downgrade when fallback would break core capabilities such as multi-image reference, precise last frame, or large-amplitude motion.