The professional generation workflow

Before any generation happens, the artist gathers references. This might mean sketching a rough composition on paper, collecting reference images from Pinterest or art books, writing out the key visual elements, or building a moodboard. The goal of this stage is to know what you are trying to make before you start trying to make it, so that when the model produces an output you can evaluate it against a clear target rather than against a vague 'does this look good.' Most amateur AI artists skip this stage entirely and start typing prompts blind. Most professional ones spend more time on this stage than on any other.

Generate dozens of low-effort outputs to find a composition and approximate look that matches your target. At this stage you are looking for a base image that is roughly right, the composition is good, the lighting is in the right ballpark, the subject is plausible, even if many details are wrong. You are not trying to produce a finished image. You are trying to find a starting point. Working artists generate 20 to 50 candidates at this stage and pick the best one or two for the next stage. The key insight is that prompts at this stage should be relatively short and simple, long prompts at this stage just add constraints that make it harder to get to a good base.

Once you have a good base image, you refine it. The two main techniques are img2imgImg2imgImage-to-image generation: starting from an existing image rather than from pure noise, and applying a prompt to modify it. (also called image-to-image) and ControlNetControlNetA conditioning technique that lets you provide structural guidance (depth, edges, pose) to a diffusion model to control spatial layout.. img2img takes your base image, adds some noise back into it, and runs the diffusion process again, but starting from your image instead of from pure noise. The 'denoisingDenoisingThe core loop of diffusion. Starting from pure noise, the model repeatedly predicts and subtracts a little noise until a coherent image emerges. strength' parameterParameterOne individual weight. Model size is quoted in parameters, in the millions or billions. More parameters can mean more capability and more cost. controls how much the new generation diverges from the original. At low denoising strength (0.2 to 0.4), the output looks almost identical to the input but with some details changed. At medium denoising strength (0.5 to 0.7), the structure is preserved but the rendering is mostly new. At high denoising strength (0.8+), the result is almost a fresh generation that only loosely respects the input.

ControlNet is the more surgical tool. Instead of preserving an image as-is, it extracts a specific structural feature from the input, a depth mapDepth mapA grayscale image encoding how far each pixel is from the camera, used as conditioning to control 3D structure., a pose skeleton, an edge map, a normal map, a segmentation mask, and uses that feature as a strict constraint on the new generation. Generating with a pose ControlNet means: 'make a new image that exactly matches this pose, but otherwise reinterpret everything.' Generating with a depth ControlNet means: 'preserve the 3D structure of the scene but change the styling and content.' ControlNet is the technique that turns image generation from 'roll the dice' into 'execute a specific creative intent.'

The key ControlNet types in 2026 are: OpenPoseOpenPoseA skeleton map of a figure's pose, used as conditioning to control the body position of generated people. (for character poses), Depth (for 3D scene structure), CannyCannyAn edge-detection map, from the Canny algorithm, used with ControlNet to make a generated image follow the outlines of a reference. edges (for line work and outlines), Normal maps (for surface orientation), Segmentation (for separating regions of the image), and Intellectual Property (IP)-Adapter (for using a reference imageReference imageAn image fed to a model as conditioning to carry a subject, style, or composition into new output, rather than describing it in words. to bias the style or identity). Working artists use 2-3 of these per image, layered together with different weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable. to produce specific effects. ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org.'s nodeNodeA single block in a visual pipeline like ComfyUI that does one operation and wires to others to form a workflow.-based interface is the dominant tool for this kind of layered ControlNet work because each ControlNet becomes a separate node in the graph and you can see exactly how they combine.

InpaintingInpaintingEditing a specific masked region of an image while preserving the rest. The classic AI editing approach. is the technique of regenerating just a specific region of an image while keeping the rest unchanged. You draw a mask over the part you want to change, write a new promptPromptThe text description you provide to a model to specify what you want it to generate. that describes what should be there, and the model regenerates only the masked region while preserving the surrounding pixels. This is the technique that fixes broken hands, bad faces, awkward poses, and unwanted background elements. It is the closest thing to 'undo this specific mistake' that the technology offers, and it is the difference between accepting whatever the model gave you and producing exactly what you wanted.

Inpainting is also where the quality difference between casual and professional users is most visible. A casual user might inpaint a single fix and call it done. A professional user might inpaint 15 different regions in sequence, first the hands, then the eyes, then the background lamp, then the texture of the wall, then a small composition adjustment in the corner. Each inpainting pass is small and targeted. The cumulative effect is an image that looks coherent because every region was fixed individually, by an artist with a clear sense of what each region needed to be.

The final stage is upscalingUpscalingIncreasing an image or video's resolution after generation, often with a model that invents plausible fine detail rather than simply stretching pixels. and color grading. Most generation happens at 1024x1024 or similar, which is too low for print or most professional uses. An upscalerUpscalerA model or tool that raises the resolution of an image or video, either faithfully recovering detail or inventing plausible new detail. model takes the 1024x1024 output and produces a 4096x4096 (or larger) version with added detail. The current best upscalers (Topaz Gigapixel AI, Magnific, Krea's upscaler) do more than just scale up, they add believable detail at the new scale, so the upscaled image looks like it was rendered at high resolutionResolutionThe pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time. rather than blown up from low resolution.

After upscaling, professional workflows typically do a color grading pass in a traditional photo editor (Photoshop, Capture One, DaVinci Resolve for video) to refine the color palette, contrast, and overall mood. The image generation model produces the content; the color grading produces the finish. Many of the images you see in professional AI artist portfolios have been through both stages, even though only the generation step gets credited.