Create an image from one sentence. You can start even without source material — describe the person, outfit, scene, and viewpoint, and get an image directly.
This skill also includes a set ofinstruction templates you can copy directly: change the model, change the pose, change the outfit — these are the common building blocks for all other skills.
Generation result examples
This skill does not require input material — only one sentence of description.
The command actually executed:
dlazy banana-pro \
--prompt 'A long-haired young Asian woman wearing a white puff-sleeve lace midi dress, sitting at a marble table inside a French-style cafe, fresh flowers and pastries arranged on the table, warm white colour grading, soft window light, front three-quarter view, waist-up framing, atmospheric editorial portrait, photorealistic, shot on 85mm, shallow depth of field, no text, no watermark.' \
--aspectRatio 3:4 --imageSize 2K \
--save docs/creative-scene/example-output.jpg
The corresponding five-slot breakdown:
| Slot | Content |
|---|---|
| Person | a long-haired young Asian woman |
| Outfit | white puff-sleeve lace midi dress |
| Scene | marble table inside a French-style cafe, fresh flowers and pastries |
| Viewpoint/Shot size | front three-quarter view, waist-up framing |
| Atmosphere/Tone | warm white colour grading, soft window light, atmospheric editorial |
Output
example-output.jpg — 3:4 / 2K, 18 credits. The person, outfit, French café scene, three-quarter front half-body shot size, and warm white tone all land according to the description.
This image can be used directly as input for subsequent targeted edits — for example, follow it with 把模特的上衣变成黑色高领修身打底衫，保持其他不变 to change the outfit.
1. Capability boundaries
| Mode | Description |
|---|---|
| Pure text-to-image | Only give a description, create an image out of thin air |
| Reference image + description | Use an image for targeted edits (change model / change pose / change outfit) |
| Vocabulary category | Description |
|---|---|
| Indoor / outdoor / home / street | Scene category |
| Half-body / full-body / close-up | Shot size category |
| Scene image | Complete scene atmosphere |
| Change model / change pose / change outfit | Targeted edit category (see section 3) |
Do not do: Do not generate portraits of specific real people; do not generate inappropriate content involving minors; do not use it to fake real product photos and usage experiences.
2. Description formula
The official reference form is:Person + outfit, scene atmosphere, image viewpoint。
一个长发小女孩穿灰色打底裤和长款羽绒服白色雪地靴，咖啡店圣诞，全身正面
└── 人物 ──┘└──────── 穿着 ────────┘ └── 场景 ──┘ └─ 视角 ─┘
Break it into five slots, and fill in whichever is missing:
| Slot | Example |
|---|---|
| Person | 长发女生 卷毛非洲小女孩 35 岁商务男性 混血模特 |
| Outfit | 白色泡泡袖蕾丝连衣裙 军绿麻花毛衣 + 米白阔腿裤 + 白色运动鞋 |
| Scene | 法式装修风的咖啡店内，鲜花布置 新疆草原，身后有草原动物 T台走秀 |
| Viewpoint/Shot size | 全身正面 半身侧面 下半身图 颈部特写 半身背面转头 |
| Atmosphere/Tone | 明亮的暖白色调，氛围感 复古色调，写真 冷色光，明亮色调 |
E-commerce-oriented experience：写真、氛围感、明亮的暖白色调 Terms like these can significantly improve the commercial usability of generated images;全身正面 / 半身侧面 Clear shot-size terms like these can greatly reduce composition rework.
3. Three types of targeted edit templates (ready to copy)
When using a reference image, use the following sentence patterns for targeted edits.Common point: they all end with "keep everything else unchanged"— this is the key to not going off track.
Change model
将主体改为一个[年龄]岁的[种族][性别][职业类型]模特，她/他有着[五官特征]、[发型]、[肤色]。
保持相同的服装、姿势和背景。
Sub-dimensions can also be changed individually:
| Dimension | Template |
|---|---|
| Skin tone | 将肤色改为[浅色/浅桃色/温暖的香槟色/蜜色/可可色]。保持原有的面部结构、种族特征、所有面部特征的大小和位置、身体姿态、体型以及光照条件完全不变。 |
| Body shape | 将身材改为[苗条/丰满/稍微丰满/肌肉型]。保持主体完全不变，只改变身材。 |
| Hairstyle | 将发型改为[双麻花辫/凌乱的丸子头/长发大波浪/低马尾/精灵短发/寸头/背头]。保持角色与背景的一致性，同时保留相同的面部特征、姿势和表情。 |
| Facial features | 将[瞳孔颜色/唇色/眉毛/睫毛/腮红]改为[具体描述]。保持角色与背景的一致性，同时保留相同的姿势和表情。 |
Change pose
将姿势改为[姿势描述]。保持完全相同的拍摄角度、面部结构、肤色和体型。
Common poses:正面站立，双臂自然下垂 / 以自然的步伐朝相机方向走来 / 侧身站立，上半身朝向相机扭转，双手放在背后以展示腰线和侧面轮廓 / 背对镜头站立，回头看向肩膀 / 像模特一样走在T台中间的步伐，一脚向前并扭动臀部 / 坐在地板上，向后倾，用双手在身后支撑身体
Change outfit
把模特的[上衣/下装/鞋子/配饰]变成[具体描述]，保持其他不变
Example:把模特的上衣变成黑色高领修身打底衫，冬季百搭款，保持其他不变 / 把模特的鞋子变成深棕色8孔马丁靴，硬朗帅气风，保持其他不变
4. Tool Invocation
This skill uses dLazy's banana-pro(a high-quality text-to-image model, optionally with reference images; excels at detailed key visuals, product images, and brand-style images, suitable for the main scenario of creating images from scratch).
How to Invoke
Two equivalent ways to write it, choose one. The unified entry point automatically selects the backend, retries on failure, creates directories and writes to disk, and estimates cost:
# A. 统一入口（推荐）：可切任意后端，加 --dry-run 不计费空跑
node scripts/gen.mjs --task creative-scene \
--prompt '<见下方 Prompt 模板>' \
--images <按下表顺序> \
--save output/creative-scene-<sku>.jpg
# B. 直接用 dLazy CLI（不想引入 Node 依赖时，效果等价）
dlazy banana-pro --prompt '...' --images ... --save output/creative-scene.jpg
Parameter Conventions (fixed usage for this skill)
| Parameter | Value | Reason |
|---|---|---|
--images |
Leave blank (pure text-to-image) / [参考图](targeted modification, up to 14 images) |
Only needed when changing model/pose/styling |
--aspectRatio |
3:4(e-commerce and recommendation vertical) / 1:1 / 9:16(short video cover) |
Mainly e-commerce vertical |
--imageSize |
1K draft exploration / 2K listing / 4K large images and print |
Use 1K during exploration to save cost |
--batch |
4(exploration phase) / 1(finalized) |
Text-to-image relies on quantity to get the best |
--save |
docs/creative-scene/output-<主题>.jpg |
Archive by theme |
English vs Chinese prompt: For scene and atmosphere descriptions, either Chinese or English works; butfor garment material and craftsmanship details, English is more stable. For targeted modification templates, Chinese is sufficient (the sentence patterns in Section 4 are proven effective).
Command Examples
# basic call: 纯文生图
dlazy banana-pro \
--prompt 'A long-haired young Asian woman wearing a white puff-sleeve lace midi dress, sitting at a marble table inside a French-style cafe, fresh flowers on the table, warm white colour grading, soft window light, front three-quarter view, waist-up framing, photorealistic, no text.' \
--aspectRatio 3:4 --imageSize 2K
# complex call: 探索期一次出 4 张 1K 草稿，定稿后再出 2K
dlazy banana-pro \
--prompt 'A long-haired young Asian woman wearing a white puff-sleeve lace midi dress, sitting at a marble table inside a French-style cafe, fresh flowers and pastries arranged on the table, warm white colour grading, soft window light, front three-quarter view, waist-up framing, atmospheric editorial portrait, photorealistic, shot on 85mm, shallow depth of field, no text, no watermark.' \
--aspectRatio 3:4 --imageSize 1K \
--batch 4 --save docs/creative-scene/draft.jpg
# 定向修改：改姿势（带参考图）
dlazy banana-pro \
--prompt '将姿势改为侧身站立，上半身朝向相机扭转，双手放在背后以展示腰线和侧面轮廓。保持完全相同的拍摄角度、面部结构、肤色和体型。' \
--images docs/creative-scene/example-output.jpg \
--aspectRatio 3:4 --imageSize 2K \
--save docs/creative-scene/output-pose-side.jpg
# 定向修改：改搭配
dlazy banana-pro \
--prompt '把模特的上衣变成黑色高领修身打底衫，冬季百搭款，保持其他不变。' \
--images docs/creative-scene/example-output.jpg \
--aspectRatio 3:4 --imageSize 2K \
--save docs/creative-scene/output-swap-top.jpg
# 先估价不真跑
dlazy banana-pro --dry-run --prompt '...' --aspectRatio 3:4 --imageSize 2K
Further Reading
| What to look up | Where to go |
|---|---|
| Authentication, multi-backend configuration, output structure, error codes | references/provider-cli.md |
banana-pro All available parameters |
references/model-flags.md |
| All options of the unified entry point | node scripts/gen.mjs --help |
5. Prompt Templates
Pure text-to-image(five-slot formula):
[人物]，穿着[穿着描述]，在[场景描述]，[视角/景别]，[氛围/影调]，
photorealistic, [镜头语言：shot on 85mm, shallow depth of field], no text, no watermark
Targeted modification(three types of templates, see Section 4):
[改模特 / 改姿势 / 改搭配 的句式]。保持[其余要素]完全不变。
Correction sentences appended based on the problem
| Problem | Append to the end of the prompt |
|---|---|
| Wrong composition (asked for full body, got half body) | Move the shot-size word earlier in the prompt, and append full-body framing, head to feet fully in frame |
| Output looks too much like an illustration | photorealistic photograph, real fabric texture, natural skin pores, no illustration, no CGI |
| Hands are mangled | hands anatomically correct, five distinct fingers, natural knuckles |
| Garment details are uncontrollable | Switch the garment description to English, and specify color / fabric / fit / neckline exactly |
| Targeted modification changed other things too | Write "keep everything else unchanged" more specifically: list item by item what must not be changed |
| A set of images has inconsistent style | Keep the atmosphere paragraph and camera-language paragraph word-for-word unchanged, only swap the person/scene |
6. Execution Flow
- Determine the mode: Creating from scratch → pure text-to-image; modifying an existing image → with reference image + Section 4 template.
- Write the description using the five slots: person / clothing / scene / perspective / atmosphere; if one is missing, fill it in.
- Exploration phase
--imageSize 1K --batch 4to quickly see the direction, lowest cost. - After finalizing,
--imageSize 2Kto produce the formal image. - For continuous edits(first set the person, then change the pose, then change the styling): use the output of the previous step as the
--imagesfor the next step, changing only one dimension per step. - Quality check: Whether the composition and shot size are correct, whether it is realistic enough, and whether the hands and clothing details are controlled.
- Images to be listedIt is recommended to run through detect-task for pre-launch inspection.
7. Common Questions
| Symptom | Cause | Solution |
|---|---|---|
| Wanted a full body but got a half body | The shot size term is placed too late | Move the shot size term earlier + append head to feet fully in frame |
| The output image looks like an illustration/CG | Lack of realism constraints | Append photorealistic + realistic texture sentence |
| Clothing details are completely uncontrollable | The description is too vague or the material is written in Chinese | Change the clothing section to English, and specify the color/fabric/cut/neckline |
| Hands are mangled | Generation randomness | Append a hand-fixing sentence + --batch 4 select images |
| Targeted edits change too much | "Keep unchanged" is written too vaguely | List item by item the elements that must not be changed |
| A set of images has an inconsistent style | The atmosphere section is changed every time | Fix the atmosphere section and the camera language section |
| Cost exceeds expectations | Always using 2K/4K for exploration | Use 1K + --batch 4for exploration, then 2K for the final version |
Tips
Visit https://dlazy.com for more information.