GPT-Image-2, FLUX Schnell, and Midjourney each handle reference images differently. Learn which strategy works best for your AI model in 2026.
By late 2026, every major AI image model supports reference images, but each model uses them fundamentally differently. GPT-Image-2 treats references as visual prompts for reinterpretation, FLUX Schnell uses them as structural guides, and Midjourney separates style from character references entirely. Understanding these differences is the difference between inconsistent output and a cohesive brand campaign.
GPT-Image-2: Reference as Visual Prompt
OpenAI's GPT-Image-2 handles reference images unlike any other model. Rather than extracting a style or copying a structure, it understands the reference — its composition, lighting, subject positioning — and reinterprets it in a new context. This makes GPT-Image-2 ideal for Hong Kong brands that need the feel of an existing asset applied to new creative directions.
For example, upload a product shot from your Spring campaign and prompt "same product in a night market setting with neon lights." GPT-Image-2 will preserve the product's positioning and lighting logic while completely changing the environment. The result is cohesive but not identical — perfect for campaigns that need variety within a consistent visual language.
GPT-Image-2 supports up to four reference images in a single generation, letting you combine compositional, lighting, and subject references simultaneously. This multi-reference capability is unique among closed-source models and gives brands unprecedented control over output consistency.
FLUX Schnell: Reference as Structure
Black Forest Labs' FLUX Schnell excels at structural preservation. When you provide a reference image, FLUX Schnell prioritizes proportions, spatial relationships, and aspect ratios above style or colour matching. For Hong Kong brands producing packaging mockups, product catalogue images, or any asset where exact dimensions matter, FLUX Schnell is the most reliable option.
A practical workflow: upload your brand's standard product shot (e.g., a Lee Kum Kee sauce bottle at a specific angle) and generate variations with different backgrounds and contexts. FLUX Schnell will keep the bottle's shape, size, and orientation nearly identical across every output, while the surrounding scene changes. This makes it invaluable for e-commerce catalogues where product consistency is non-negotiable.
FLUX Schnell runs fast enough for rapid iteration — you can generate 20-30 structural variations in the time other models produce four. Use this speed to test multiple backgrounds, lighting scenarios, and product arrangements before committing to post-production.
Midjourney: Separate Style and Character References
Midjourney's dual-reference system — --sref for style and --cref for character — gives creators granular control unmatched by any competing platform. Style references capture colour palettes, textures, and aesthetic qualities. Character references preserve facial features, body types, and clothing across generations. Using both simultaneously lets you lock in a brand's visual identity and a recurring subject in a single generation.
For Hong Kong agencies running multi-image campaigns with the same talent, --cref is transformative. Upload one headshot of your brand ambassador, and Midjourney can place them in dozens of scenes while keeping their face, hair, and outfit consistent. Pair with --sref from your brand style guide, and every output belongs to the same campaign visually.
The separation of concerns also means you can mix and match: use one brand's style reference with a different character reference for concept testing, or keep the same character while exploring entirely new visual directions. This flexibility is why Midjourney remains a staple in Hong Kong creative agency workflows.
Nano Banana 2: Speed and Mood Matching
ByteDance's Nano Banana 2 prioritizes speed and emotional tone over structural precision. It reads the colour story and mood of a reference image — warm, cool, dramatic, airy — and applies that feeling to new outputs. While it may not preserve exact product dimensions, it excels at social media content where vibe consistency matters more than pixel-perfect positioning.
Hong Kong brands running daily social content can use Nano Banana 2 to maintain a consistent Instagram feed aesthetic without generating every post from scratch. Upload one on-brand lifestyle photo as reference, and the model will match the colour palette and lighting mood across product shots, lifestyle scenes, and text overlays.
ComfyUI and Open-Source: Maximum Customization
For studios that need total control, ComfyUI workflows with Stable Diffusion offer the most flexible reference pipelines available. IP-Adapter handles style and content reference independently. ControlNet provides structural guidance via edge maps, depth maps, and pose skeletons. Reference-Only techniques make the model treat your reference as a starting point for diffusion.
Hong Kong production studios running custom ComfyUI pipelines can dial in reference adherence from 0% (pure text-to-image) to 100% (near-exact reproduction). This level of control is overkill for most quick-turnaround work, but essential for high-budget campaigns where brand guidelines are strict and deviations aren't tolerated.
Choosing the Right Reference Strategy
Match your reference approach to your primary output goal. For structural consistency — product shots, packaging, catalogue images — use FLUX Schnell or a ControlNet ComfyUI pipeline. For creative reinterpretation with brand DNA — campaign assets, social content, ad variations — GPT-Image-2 is the best fit. For campaigns needing both a consistent character AND a consistent style — talent-driven ads, mascot campaigns, testimonial series — Midjourney's dual-reference system is the clear winner. For high-volume social content where speed and mood consistency matter most, Nano Banana 2 delivers reliable results faster than any alternative.
Frequently Asked Questions
Q: Can I use multiple reference images in one AI generation? A: Yes. GPT-Image-2 accepts up to four reference images simultaneously. Midjourney lets you combine --sref and --cref independently for multi-reference control.
Q: Which model preserves product packaging layouts best from a reference? A: FLUX Schnell offers the strongest structural preservation. ControlNet-based Stable Diffusion workflows are comparable but require more setup time.
Q: Does reference image resolution affect output quality? A: Yes. All models interpret high-resolution, well-lit references more accurately. Compressed or low-resolution references produce muddier results with less fidelity.
Q: Can I use my brand logo as a reference image? A: Most models will reinterpret rather than exactly replicate logos. For precise logo placement, generate the base image first and add the logo in post-production.
Q: What reference strength setting should I start with? A: Start at 0.7. Use 0.8-0.9 for strict brand consistency, or 0.4-0.6 for creative exploration while maintaining brand feel.
Q: Do Hong Kong brands use different reference strategies for Chinese vs English campaigns? A: The reference technique is identical, but brands often maintain separate reference image sets for each language market due to different visual tone preferences in Chinese vs English-language audiences.
Q: Are there copyright concerns with reference images in AI generation? A: Use only images your brand owns or has licensed. Most model terms permit commercial use of generated content, but the reference image itself must be your IP.