Seedream: how to write prompts the model actually understands

What every Seedream version shares

The Seedream line shares one architectural idea — hierarchical prioritization. The model reads the prompt top to bottom and assigns the highest weight to whatever is mentioned first. The main subject always goes first. A subject buried at the end of a paragraph loses priority, and the framing shifts to the scene or lighting instead.

This hierarchy has five levels: Subject → Style → Composition → Lighting & Atmosphere → Technical details. Each successive version (4.0 → 4.5 → 5 Lite) expands the style range, improves in-image text rendering and handling of complex scenes, but the underlying hierarchy stays the same.

All versions perform better in English than in any other language. On 4.5 and 5 Chinese is also supported, but English remains the stable reference for production prompts.

• 5-level hierarchy: Subject → Style → Composition → Lighting → Technical
• Main subject always goes at the start of the prompt
• Optimal length 30–100 words
• English yields the most stable results
• Aspect ratio via --ar

Prompt structure

Canonical formula: `[Subject] + [Style] + [Composition] + [Lighting/Atmosphere] + [Technical parameters]`.

Subject — concrete, not abstract. «A fluffy orange tabby cat sitting on a windowsill» instead of «a cat». Style — photographic («portrait photography», «product photography»), cinematic («cinematic», «film noir»), or art style («oil painting», «watercolor»).

Composition — angle, framing, layout («close-up», «rule of thirds», «overhead perspective»). Lighting — a specific scheme («golden hour», «studio lighting», «Rembrandt lighting»). Technical — lens, depth of field, format («85mm lens», «shallow depth of field», «--ar 16:9»).

Not every level needs to be filled. Minimum — Subject + Style + 2–3 descriptors. Maximum — 30–100 words; past that the model starts losing focus.

Six base prompt types

Portrait photography — Subject + appearance + portrait photography + 85mm lens + lighting + background. Realistic skin and face rendering is one of Seedream's strengths.

Product photography — product + material + clean background + commercial style + studio lighting + ultra-sharp. The key is clean background and precise light.

Cinematic scenes — scene + character + cinematic + 35mm lens + dramatic lighting + atmosphere. Cool palettes, rim light, rain, and neon work best.

Concept art and illustration — subject/world + concept art / digital painting + epic scale + volumetric light + intricate details. Use the words «epic», «volumetric», «intricate» — Seedream reacts to them strongly.

Stylized images — subject + a specific style (watercolor, oil painting, anime) + color palette + mood. Posters and typography — poster/banner + text «...» in quotes + font style + format.

Lighting and atmosphere

Lighting is the most sensitive zone in Seedream. A specific light-scheme name yields a radically better result than a generic «good lighting».

Classic schemes: «Rembrandt lighting» (triangular highlight under the eye), «butterfly lighting» (butterfly shadow under the nose), «split lighting» (half the face in shadow), «loop lighting» (soft nose loop). These are photography terms — the model knows them.

Time of day: «golden hour» (warm sunset), «blue hour» (cool post-sunset sky), «overcast» (diffused cloudy), «noon» (hard midday). Source: «studio lighting», «natural light», «candlelight», «neon glow», «moonlight».

Quality: «soft diffused light», «dramatic side lighting», «high-key» (bright, commercial), «low-key» (dark, noir), «volumetric light» (visible beams).

Common mistakes

1. 1. Subject not at the start of the prompt

   Hierarchical prioritization assigns the highest weight to whatever is mentioned first. A prompt that begins with «Beautiful golden hour lighting in a forest with a cat» makes the forest or the lighting the main subject, not the cat. The correct form is «A cat in a forest with golden hour lighting».

2. 2. Negatives in the main text

   «No watermark, no text» in the main prompt is read by Seedream as scene description — possibly even adding a watermark. All bans go into the platform's separate negative_prompt field. When that field isn't available, phrase positively: «no cluttered» → «clean background».

3. 3. Conflicting styles

   «Photorealistic oil painting watercolor cartoon» — the model has no idea which game to play. Pick one dominant style plus at most one secondary modifier. «Photorealistic with subtle film grain» works; «realistic cartoon sketch» does not.

4. 4. Adjective overload

   Twenty weak descriptors («beautiful, nice, gorgeous, stunning, amazing, breathtaking») work worse than 3–5 strong concrete ones («Rembrandt lighting, 85mm lens, shallow depth of field, golden hour, weathered leather»). Specifics always beat generic admiration words.

5. 5. No style specified

   Without an explicit style Seedream picks at random: photorealism one time, illustration the next, 3D render after that. Style is a required element. Minimum: «photorealistic», «cinematic», «watercolor», «oil painting» — one of these has to be in every prompt.

Before / after examples

beautiful girl with flowers

A young woman holding a bouquet of pale pink peonies, photorealistic portrait style, 85mm lens, shallow depth of field, soft natural window light from the left, neutral linen background, gentle expression, subtle warm tones, --ar 4:5.

product photo

A matte black ceramic coffee mug on a white seamless background, commercial product photography, soft three-point studio lighting, ultra-sharp details, slight overhead angle, clean composition, subtle shadow on the right, --ar 1:1.

movie poster about space with title beyond the stars

Movie poster for a sci-fi epic, text "BEYOND THE STARS" in bold metallic sans-serif typography centered at top, dark deep-space background with a single illuminated planet on the right, dramatic cinematic composition, cold blue and white color palette, subtle film grain, --ar 2:3.