Prompting at a senior level

A well-constructed promptPromptThe text description you provide to a model to specify what you want it to generate. for a frontier model in 2026 has roughly five components, in roughly this order: subject, action, setting, style/medium, and technical specifications. Each component is a complete phrase, not a list of tags. Here is an example:

'A young woman in her thirties with shoulder-length dark brown hair, sitting alone at a small wooden table in a quiet European cafe, reading a paperback book and absent-mindedly stirring her coffee with her left hand. Late afternoon sunlight streams through tall arched windows behind her, creating long shadows across the table. Shot on a 50mm prime lens at f/2.8, shallow depth of field with the background slightly blurred, color palette of warm amber and deep brown with cool blue shadows.'

That prompt is 90 words. It would have been considered absurdly long in 2023. In 2026, it is the right length for a frontier model that is built around a strong text encoderText encoderA neural network that converts text prompts into numerical embeddings that the diffusion model can use. Modern models use combinations including CLIP and T5.. Each phrase contributes specific information that the model can act on. The subject is described physically. The action is specific (reading, stirring with the left hand) which gives the model something to anchor the pose. The setting is concrete enough to guide composition. The style component specifies a real photographic look. The technical specification (50mm, f/2.8) tells the model what kind of camera optics to render.

The mistake most people make is to write prompts that are either too short to give the model enough to work with, or too long and full of conflicting instructions. The fix is to be specific without being contradictory, every phrase should add new information rather than restating something already said.

Negative prompts let you tell the model what to exclude from the output. They were essential in the SD 1.5 era because the base models had so many failure modes that you needed to actively suppress them ('blurry, low quality, bad anatomy, extra fingers, watermark, signature, jpeg artifacts'). They are much less necessary with modern models, but they still have specific uses.

The right use of negative prompts in 2026 is to suppress specific visual elements that you do not want and that the model is likely to add anyway. If you are generating outdoor scenes and the model keeps including birds you do not want, 'birds' in the negative promptNegative promptA description of things you do not want in the output. The model is steered away from these concepts during sampling. will suppress them. If you want a portrait without glasses and the model keeps adding glasses, 'glasses, eyewear' in the negative prompt will fix it. The wrong use of negative prompts is to dump in long lists of generic quality complaints. With FLUX and MM-DiTMM-DiT (Multimodal Diffusion Transformer)A variant of DiT that processes text and image (or audio) tokens in separate parallel streams joined by attention. Used by Stable Diffusion 3, FLUX, and most newer models. Improves prompt following significantly. models, those long negative prompt lists actually hurt output quality by confusing the model about what you want.

Every generation starts from a tensor of random noise. The 'seed' is the number that initializes the random number generator that produces that noise. If you use the same seed and the same promptPromptThe text description you provide to a model to specify what you want it to generate. and the same model, you get the same image. If you change the seed and keep everything else the same, you get a different image, same prompt, same intent, different specific realization.

Most people treat the seed as random and ignore it. Working artists treat the seed as a creative variable. Once you find a seed that produces a composition you like, you can hold it constant while you iterate on the prompt. You can swap one element of the prompt and see how that single change affects the output, without the random variation drowning the signal. You can use the same seed across multiple generations to produce a series that has a coherent visual identity. Seed control is a small thing but it matters enormously for production work where you need consistency across generations.

The samplerSamplerThe algorithm used to perform the denoising steps. Different samplers (Euler, DPM++, etc.) have different speed-quality trade-offs. is the algorithm that runs the denoisingDenoisingThe core loop of diffusion. Starting from pure noise, the model repeatedly predicts and subtracts a little noise until a coherent image emerges. process, there are several to choose from, and they produce subtly different results. The step count is how many denoising iterations the sampler runs. More steps means more refinement but also more time. The combinations are not infinite but they are numerous, and there is no single right answer.

The pragmatic guidance is: for FLUX, use Euler with 28-32 steps for normal generation, or 4-8 steps for FLUX schnell (the fast variant). For SDXL, use DPM++ 2M Karras with 25-30 steps. For SD 1.5, use DPM++ 2M Karras with 20-25 steps. The difference between samplers is real but small, switching samplers usually changes output by 5-10 percent in subjective quality, while switching seeds or prompts changes it by 50 percent or more. Do not spend energy on sampler choice until you have already nailed prompting and seeds.