Generative AI: The 2026 Field Guide

So far we have covered the big general-purpose models, the ones that take a text promptPromptThe text description you provide to a model to specify what you want it to generate. and produce an image or a video. But the production pipeline for any serious creative work involves a lot more than just generation. You need to edit, refine, fix, animate, upscale, and combine. There is a whole category of specialized models that handle these specific tasks, and many of them are accessible through Krea, Flora, and similar platforms. We will walk through the main ones.

Editing is structurally a different problem from generation. In generation, the model starts with random noise and produces an image from scratch. In editing, the model starts with an existing image and applies a targeted change while preserving everything else. This requires a different kind of training and a different set of architectural considerations.

One shift is worth flagging because it changes how these tools are used, not just which is best. Instruction-based editingInstruction-based editingEditing an image by giving a text instruction, such as change the background to a beach, and letting the model decide what to change and what to leave alone, without a hand-drawn mask. has become chainable: rather than one edit and done, creators now run a sequence, remove the sign, then warm the light, then add a reflection, each instruction building on the last result, treating the model as a collaborator that holds the image steady between changes. Kontext, Nano Banana, Qwen-Image-Edit, and Runway's Aleph for video all support this conversational loop, and it is covered as a workflow in the craft chapter. The takeaway for this section is that the editing models are increasingly judged not only on a single edit's quality but on how gracefully they let you stack edits without the image drifting.

The classic approach: inpainting and masking

The original way to do AI image editing was inpaintingInpaintingEditing a specific masked region of an image while preserving the rest. The classic AI editing approach.. You marked a region of the image with a mask (literally painting over it in a separate alphaAlpha (LoRA alpha)A scaling factor that controls how strongly a LoRA's learned change is applied, effectively its volume knob, tuned separately from rank. channel), and the model regenerated only that region while keeping the rest unchanged. This worked, but it required the user to do the masking manually, and the regenerated region often did not blend smoothly with the surroundings. Most early Stable Diffusion workflows used inpainting for tasks like 'remove this object' or 'change this part of the image,' and the results were workable but tedious.

The modern approach: instruction-based editing

The current generation of editing models works very differently. Instead of requiring a mask, they take an image and a text instruction ('change the background to a beach,' 'put the person in a red jacket') and apply the change in-place. The model decides what to change and what to leave alone based on the instruction, using what is called 'in-context conditioningConditioningAny extra input that steers generation beyond the prompt, such as a depth map, pose, edge map, or reference image.', the input image is passed to the model as additional context, and the model is trained on enormous numbers of (original, instruction, edited) triplets so it has learned what kinds of changes correspond to what kinds of instructions.

The major instruction-based editingInstruction-based editingEditing an image by giving a text instruction, such as change the background to a beach, and letting the model decide what to change and what to leave alone, without a hand-drawn mask. models in the current generation are:

FLUX Kontext (Black Forest Labs)

Already covered in the image-models part. The most widely used open-weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable. editing model. Kontext [pro], [max], and [dev] variants. Strong identity preservation, fast iteration, and a permissive license for the [dev] version. This is the editing model that the open-source community has standardized on for most workflows.

Nano Banana Pro (Google)

Already covered in the image-models part. Excellent for text-heavy edits and infographic-style work because of its ability to render legible text. Particularly strong at multi-image composition, you can give it 14 reference images and a description, and it will produce a coherent composite.

Qwen-Image-Edit (Alibaba)

The editing-specialized variant of Qwen-Image. 20 billion parameters. Strong on text-driven semantic edits, style transfers, and bilingual text rewriting in existing images. Open-weights, which is unusual for a model of this caliber. This is the editing model that Chinese-language users tend to default to, because of its superior handling of Chinese text in images.

Seedream Edit (ByteDance)

The editing variant in ByteDance's Seedream line. Focused on commercial design use cases, product photography refinement, marketing image variations, layout adjustments. Available through ByteDance's Application Programming Interface (APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface.).

Runway Aleph

Already covered in the video-models part. Aleph is the video equivalent of these image editing models, instruction-based video editing where you can ask for changes to existing footage. The integration with Runway's broader creative suite makes it especially useful for post-production workflows.

The current instruction-editing models

Google set the tone with Gemini 2.5 Flash Image, released on August 26, 2025 and nicknamed Nano Banana. It blends multiple reference images into one, holds a character's appearance steady across successive edits, and takes targeted natural-language instructions such as background changes, object removal, and pose adjustment. Crucially it works turn by turn, so you refine an output conversationally rather than re-prompting from scratch. Google later extended the line with Nano Banana Pro, built on Gemini 3 Pro Image, which added stronger reasoning and search grounding.

The open and challenger side moved just as fast. Alibaba released Qwen-Image-Edit on its 20-billion-parameterParameterOne individual weight. Model size is quoted in parameters, in the millions or billions. More parameters can mean more capability and more cost. Qwen-Image backbone, with a dual path: a vision-language model reads what the scene means while a variational autoencoder preserves pixel-level detail, which lets it handle both high-level semantic edits and precise appearance edits, including editing text rendered inside the image in Chinese and English. Black Forest Labs released FLUX.1 Kontext, a flow-matching family (pro and max, plus an open-weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable. 12B dev model) built for fast iterative editing with strong character consistencyCharacter consistencyThe problem of generating the same character, with the same face and identity, across many images, poses, and scenes. It is the main bottleneck holding back AI narrative work. and local edits from text alone. ByteDance covered both ends of its stack: SeedEdit focused on portrait retouching, background swaps, and lighting, and the later Seedream paired a text-to-image model with a dedicated edit model that accepts several reference images and is known for accurate in-image text.

Across all of these, three capabilities recur and explain why editing now anchors real workflows. First, multi-turn conversational editingConversational editingRefining an image through a sequence of instructions where each builds on the last result, treating the model as a collaborator that holds the image steady between changes. Also called multi-turn editing., where each instruction builds on the last, so a session becomes a dialogue rather than a lottery. Second, inpaintingInpaintingEditing a specific masked region of an image while preserving the rest. The classic AI editing approach. and outpaintingOutpaintingExtending an image beyond its original borders by generating new content that fits with the existing edges., filling or replacing a masked region or extending an image beyond its original borders. Third, consistent-character editing, keeping a recognizable person or product stable across many images, the feature that makes these tools usable for storyboards, ad variants, and catalogs rather than one-off art.

Image and video editing models