A prompt is an instruction interpreted by a particular model and workflow. Start with its documentation and a small comparison, not a universal formula. These original examples are suggested exercises; they are not newly generated or benchmarked results.
The same intention, three formats
Goal: a green mug on a desk, softly lit from the left. Keep the subject consistent while changing the format you test.
Stable Diffusion 1.5 / SDXL — a phrase-based starting point
green ceramic mug on a dark wooden desk, soft window light from the left, close-up product photograph, shallow depth of fieldFLUX.2 dev / Z-Image Turbo — a connected scene description
A green ceramic mug sits near the front edge of a dark wooden desk. Soft window light enters from the left, making a gentle highlight on the mug. Photograph it at tabletop height, with the background softly out of focus.Video — add action and camera timing
Begin with the green mug centered on the desk. During the shot, the camera slowly moves closer while a small curl of steam rises from the mug. Keep the window light steady and finish without a cut.Stable Diffusion is a family, not one prompt recipe
SD1.5, SDXL and their fine-tuned checkpoints can respond differently. Some checkpoints were tuned around tags; others handle prose well. Check the model card for recommended wording, resolution and conditioning. Short phrases separated by commas are easy to edit and compare, but commas themselves do not guarantee better images.
In a compatible ComfyUI text encoder, weighting such as (soft window light:1.2) can emphasize a phrase. Parser support matters: copying that syntax into another model or interface may change its meaning or do nothing. Try modest changes and inspect the output.
Optional negative example — only with supported conditioning
blurry, duplicated objects, distorted handle, watermarkA negative prompt is a conditioning input, not a list of things the model can never produce. Avoid a huge inherited list that you have never tested.
FLUX and Z-Image: describe relationships
Use natural-language descriptions as a starting point. Name the subject, its position, the important action and the light. “A person holds the mug beside the window” is clearer than a disconnected list if the relationship matters. Add text you want rendered with explicit wording, then inspect the spelling in the result.
Do not assume SDXL negative prompts, CFG values or weighting syntax apply unchanged. Z-Image Turbo’s CFG 1 baseline is a particularly poor place to infer negative-prompt support just because a box exists. See the matching Z-Image and FLUX.2 dev guides.
Video: one understandable shot
| Part | Write it concretely |
|---|---|
| Subject and starting state | Who or what is visible at the beginning? |
| Action | What changes during the shot? Start with one action. |
| Camera | Static, slow push-in, pan or another deliberate move? |
| Timing | What happens first, next and at the end? Avoid impossible timing. |
| Sound / speech | What is heard? Identify the speaker and exact words when needed. |
| Continuity | Which identity, clothing, environment and lighting should stay consistent? |
A controlled prompting exercise
- Choose a model, exact workflow and fixed seed. Keep the resolution, sampler and guidance unchanged.
- Write the shortest prompt that expresses the desired scene. Generate a baseline.
- Add one relationship or lighting instruction. Compare subject placement, lighting and unwanted changes.
- Repeat across a few seeds before deciding the wording is reliable. Save the prompt with the model and settings, not by itself.
Copy, adapt, inspect
Copy buttons are a convenience, not a claim that these prompts are optimized for every checkpoint. Treat each example as a starting hypothesis. The tutorials show which configurations were demonstrated and which instructions are suggestions.