Image Generation Best Practices

Learn prompting technique and best practices to make best use of image generation models.

Quickstart

Think like an art director, not just a describer: specify what you want to see in the image, including composition, mood, and technical style.
Write what you want to appear: use direct statements (e.g., “a watercolor painting of a red fox in a snowy forest”) rather than commands or questions.
Focus each prompt on a single subject or concept per image. For multi-image workflows (e.g., style transfer, inpainting), clarify the relationship between each input image and the desired output.
Integrate photographic or artistic terms to guide the model’s rendering—mention lens style, lighting, atmosphere, or artistic movement if relevant.
Iteratively refine prompts: start simple and expand details step by step if the first output isn’t close enough.
If image text is crucial, enclose desired text in quotes within your prompt.

What Gives Best Results

Strategy	Why It Helps	Tips
Be specific & concrete	Stops ambiguous/unstable outputs	“sunset over misty mountains with pines silhouetted in foreground”
One main subject/focus	Models can be easily confused	Stick to “child reading a book on a park bench” rather than several actions/subjects at once
Use technical/art terms	Model maps these to visual cues	“macro shot, soft light, shallow depth of field, watercolor style”
Guide composition/camera	Controls framing and realism	“over-the-shoulder perspective, wide shot, centered subject, Dutch angle”
Define lighting/atmosphere	Increases depth and richness	“backlit by golden hour sun, morning mist, subtle haze, lens flare”
Specify color & style	Controls consistency and mood	“pastel palette, cinematic color grading, cool tones, minimalist”
Iterative prompt refinement	Gets closer to your vision	Adjust one detail at a time, test and tweak composition, colors, or subject

What to Avoid

Request/conversation style: avoid “Can you please draw…” or “Please generate…”. Use direct descriptions.
Overloading one prompt: don’t cram multiple unrelated objects, actions, or styles into a single prompt.
Too abstract or metaphorical: e.g., “emotions in color” is too vague—be visual and literal.
Long chains of unrelated adjectives: “epic, huge, cold, warm, soft, electric, ancient, modern, winter” is counterproductive.
Omitting negative constraints: expect more distortions without them, especially from generalist models.
Rapid mood/style switching: don’t mix “pixel art fantasy” with “photo-realistic noir” in one prompt.
Ignoring composition/camera guidance: model defaults can be bland or off-balance.

Examples

Example 1 – Fox in Forest

Prompt:


#Details - A vibrant, hand-painted illustration of a red fox standing alert in a snowy forest; soft blue morning light glinting off the snow.
 
#Composition - Centered, eye level, wide shot with tall pines in background.
 
#Lighting & atmosphere:-  Subtle mist, illuminated breath from the fox, cool color palette with warm highlights.

Output

Image To VideoExample

Example 2 – Product Photography Mockup

Prompt:


#Details - Modern, minimalist photo of a white ceramic coffee mug on a wooden table; soft sunlight streaming from left; shallow depth of field focuses on the mug, gentle bokeh in background.
 
#Composition - Medium close-up, mug centered.
 
#Lighting & effects- Morning light, crisp highlights, natural shadows.

Output

Image To VideoExample

Compare to Poor Prompts

“Make a great image of nature.”
“Beautiful scene with animals, trees, sun, rain, clouds.”
“Please draw a Van Gogh version of a mug with sparkles and a cat and words on it.”

Technical / Artistic Terms Cheat Sheet

Category	Example Terms
Camera / Framing	overhead, close-up, wide shot, Dutch angle, bokeh
Motion / Blur	soft focus, shallow depth of field, motion blur
Lighting	rim light, soft light, volumetric, studio lighting
Color / Style	cinematic grade, pastel palette, monochrome, sepia
Painting Styles	watercolor, oil paint, digital painting, sketch
Negative Prompts	—no watermark, —no text, —no distortions
Quality	ultra-detailed, 8K, hyper-realistic, photorealistic