Best AI image editing models with reference images

The key question is what must stay unchanged: a person’s face, a product’s shape, the background, or the visual style. Those requirements call for different controls. A reference image guides a model; it does not guarantee that every logo, facial feature or label survives generation.

Magic Hour publishes this guide and includes its own product in the comparison. Treat the recommendations as editorial guidance from a vendor, and verify the linked first-party product details and your own output requirements before choosing a tool.

Best reference-image editing options compared

This guide was reviewed on September 13, 2026. It compares documented capabilities and practical selection criteria. Magic Hour publishes it and is included; the table does not assign invented image-quality scores.

For a separate quantitative view, use the dated AI image editing model leaderboard. It combines several public benchmark sources; use this guide for workflow fit and the leaderboard for source-by-source model scores.

Choose the reference control that matches the job

For example, changing the background behind a bottle is different from creating a new scene that happens to contain a similar bottle. If the actual packaging must remain exact, preserve the product cutout and build the scene around it.

1. Magic Hour: a practical browser starting point

Magic Hour’s image editor takes an uploaded image and a description of the change. Its product page lists editing options and models, including Qwen Edit, GPT Image 2, Nano Banana models and FLUX.2 Klein. Check the current selector for the model and settings available to your account.

The public tool offers three guest edits per day without signup and watermark-free image downloads. Free outputs are for personal use; paid plans permit commercial use. A free browser edit does not establish the same model access, reference count or allowance as a paid or API workflow.

Try a precise instruction: “Remove the chair on the left. Keep the person, camera angle, wall color and remaining furniture unchanged.” Compare the output with the original, especially around the edit boundary.

Choose Magic Hour when the workflow may continue into image-to-video after the image is approved. For an integration, use the model reference rather than assuming every available model accepts the same inputs.

2. OpenAI GPT Image: conversational and instruction-based edits

OpenAI introduced ChatGPT Images 2.5 on September 8, 2026. ChatGPT adds sketch, template and image-comment workflows, while the API offers GPT-Image-2.5 Flare and Sunburst. These are options to evaluate for focused changes and successive revisions.

The image-generation guide documents image editing, image inputs and masking. Confirm the selected model’s supported controls and dimensions. A ChatGPT subscription does not pay for separate API requests.

Describe the requested change and the features to preserve. If you are combining images, explain each image’s role: source subject, target scene, style or composition. Avoid treating a reference upload as an implicit instruction to copy every element in it.

OpenAI’s launch describes improved editing fidelity, but that is a provider claim, not evidence that every product label or person will remain exact in your files. Inspect the details that matter to the final use.

3. Nano Banana: match the exact model to the references

Google’s Gemini image guide distinguishes Nano Banana 2, Nano Banana Pro, Nano Banana 2 Lite and the original Nano Banana. Their reference capabilities differ.

The current guide permits up to 14 reference images across the Gemini 3 image models, with different categories of supported references. It describes Nano Banana 2, or Gemini 3.1 Flash Image, with up to 10 object references and four character references. Nano Banana Pro, or Gemini 3 Pro Image, lists up to six object references, five character references and three style references within the 14-image total. Do not transfer those figures to every Nano Banana interface or variant.

Use this family when you need to specify several source elements. Name their roles clearly: “Use the person from image 1, the jacket from image 2 and the setting from image 3.” Start with only the references needed to remove ambiguity; extra images can introduce conflicting cues.

For a single product, one clear view plus a close view of an important detail can be more useful than a large set of unrelated inspiration images. The downloaded result still needs inspection.

4. FLUX.2: explicit multi-reference composition

Black Forest Labs’ FLUX.2 editing documentation describes text-guided edits and combining multiple source images. It distinguishes up to eight references through the documented API workflow from up to ten in the playground, with variant-specific capabilities.

The family includes max, pro, flex and klein options. Choose the specific model and host before comparing costs or assuming the same controls are available. A capability in Black Forest Labs’ playground is not automatically exposed by every app that offers FLUX.

A useful brief assigns one role to each source: “Keep the room layout from image 1. Replace only the sofa with the sofa in image 2. Match the room’s existing light and perspective.” Review the sofa’s proportions and the surrounding floor after the edit.

This is a practical option for multi-source scene building. It should not be described as universally better or worse at identity preservation without a comparison using the same inputs and requirements.

5. MAI-Image-2.6: another current model to evaluate

Microsoft made MAI-Image-2.6 and the faster-oriented Flash variant available in Microsoft Foundry public preview on September 4, 2026. The announcement describes multi-image reference editing, web grounding and dynamic aspect ratios.

The Foundry documentation provides the actual generation and edit endpoints, input requirements, output format and deployment quotas. Follow those endpoint details rather than assuming a feature’s announcement describes every available client or setting.

Include MAI-Image-2.6 in a model evaluation if you are choosing an API for reference edits. Preview availability and output limits matter alongside visual quality. There is no reason to choose it solely because a general leaderboard places it ahead by a few points.

What the independent benchmark does—and does not—show

The Artificial Analysis editing leaderboard ranks results through blind user preferences for edits to the same input and instruction. In the September 10 snapshot, MAI-Image-2.6 scored 1122 and GPT Image 2 at high quality scored 1117, with overlapping reported 95% confidence intervals. That supports including both in a shortlist; it does not establish a decisive quality difference or perfect reference fidelity.

That GPT Image 2 result is not an evaluation of the newly released GPT Image 2.5 models. Aggregate preference also does not tell you whether a particular bottle label, facial feature or catalog requirement will pass review.

6. Midjourney: use its current editing workflow

Midjourney’s August 27 Edit Model announcement introduced a V8.2 editing workflow with instruction-based edits, up to four image references, inpainting and outpainting. It also describes personalization, moodboards and style-reference support.

The current web Editor supports uploaded images, selections and layers. This makes older “Discord-only” or “style reference only” descriptions misleading. Older Omni Reference instructions are not a substitute for checking the current Edit Model workflow.

Choose it when you want to combine art direction with direct image edits. Use a selection for a local change and a reference for the specific visual information you want to introduce. Review the result rather than assuming that reference weight alone locks identity.

7. Photoshop: select the area, then specify the reference

Adobe’s Reference Image instructions give a concrete workflow: create a selection, choose Generative Fill, add a reference image and select Firefly Fill & Expand.

The controls distinguish an Object reference from a Whole image reference, and swapping the selected area from placing an object into it. Those choices help clarify whether you want to replace something or add a new item while retaining the setting.

Use Photoshop when selections, layers and manual repair are essential to the deliverable. For exact packaging, keep the original label or product layer available instead of repeatedly asking generation to recreate it. Adobe’s partner-model workflows have their own controls, so do not assume every Firefly option behaves identically.

8. Stable Diffusion with adapters: a custom control pipeline

For technical users, a compatible Stable Diffusion pipeline can combine image conditioning, structural controls and inpainting. Hugging Face’s IP-Adapter documentation explains image-conditioned generation and combinations with ControlNet.

IP-Adapter uses image information to guide a result; ControlNet can add a structural condition such as pose, edges or depth. The exact combination depends on the base model, adapter, pipeline and weights. An adapter built for one model family is not automatically compatible with another.

Choose this route when the control requirement justifies setup and maintenance. Keep the model versions and settings with the project. Self-hosting also requires suitable hardware and applicable model licenses; it is not automatically the cheapest way to finish a small batch.

A reference-editing brief you can reuse

Use this structure to make the task clear:

Source: which image is the starting photograph? References: what should each additional image contribute? Change: what is the one main edit? Preserve: which details must remain? Output: what dimensions, format and intended use are required?

This is an example instruction, not a tested output or a guarantee. If a logo must be exact, retain it from the source image or an approved graphic asset.

Check the result before making a batch

Judge the result at the size where it will be used, and zoom in on important details. If you compare models, use the same source, instruction and acceptance criteria, then record billable attempts and correction time. Our image-generation pricing guide explains how to compare the cost of accepted outputs.

For repeatable Seedream prompts and a practical workflow, use the Seedream editing guide.

Separate preservation from transformation

Direct answer: The best reference-image editor changes the requested region while preserving everything else. Evaluate instruction following and collateral damage separately.

Five-edit benchmark. Change an object, background, color, pose detail and text-adjacent region. Keep source images, prompts and output size fixed. Use portraits, products and complex scenes.

Review. Score requested change, identity and geometry preservation, text and logo integrity, edge quality, color consistency and unexpected edits outside the target area.

Decision rule. Use precise editors for localized production work. Use more generative systems when reinterpretation is acceptable. One model can excel at creative transformation while failing strict preservation.

Try the workflow: Open the matching Magic Hour tool. Use the same representative input and acceptance criteria before comparing results.

Continue learning

Continue learning: how AI image generation works, best AI image generators, and best AI image editors.