Image creation

The image page is a two-column layout: the conversation list on the left, and the conversation with the composer and results on the right.

Describe what you want

Write the subject, setting, composition, lighting, and style in the composer at the bottom. The submit button states how many credits the run will cost; while a run is active the button turns into a stop action ("停止生成", or "取消排队" while queued).

References

References come from three places: uploading a local image, picking an existing item from "资产库", or choosing one from "最近上传".

• The page supports several references at once. The ceiling comes from the channel (9 for channels using the New API call format); going over fails at submit.
• References are ordered: state their roles in the description, for example "使用图片 1 的人物身份" or "参考图片 2 的色彩关系". Check the order before submitting.
• "加入参考图" on a result card turns the previous output into a reference for the next round.
• The "资产库" dialog only accepts text and image assets; other kinds are rejected with "生图工作台只能使用文本或图片素材".

Image settings

The settings are remembered in the browser, so a new conversation starts from your last choices.

Prompt optimization

"AI优化提示词" asks an Agent to expand your description with concrete visual detail; the result is written back into the composer. It uses the default text model, and failures report "提示词优化失败".

Running and results

• Status moves from 排队中 to 生成中 to 已完成. Multi-image runs report "已完成 N 张，失败 M 张", and failed cards carry the reason.
• Result card actions: 拷贝链接, 上传到云储存 (then 复制云储存地址), 添加到素材, 加入参考图, 下载, and 重新生成.
• "添加到素材" saves the result into My assets with the source "生图工作台". To reuse it on the canvas, save it to assets first and then insert it from the canvas toolbar's "我的资产".
• A run that waits longer than 5 minutes is marked as failed (the administrator can tune AI_TASK_WAIT_TIMEOUT_SECONDS).

Conversation management

Every conversation in the left list is stored on the server, so it follows you across devices. You can "修改标题" or "删除记录"; deleting is blocked while a run is in progress. A conversation keeps the description and results of every round, so you can continue from the previous one.

Example: Keep the person's identity and clothing, change the setting to a rainy night street, and use cool cinematic lighting.

Example: Use image 1 for the person's identity and image 2 for the color relationship, then re-frame it as a 16:9 poster.

Troubleshooting

• "请联系管理员配置默认生图模型": no default image model has been chosen, or the channel has no model marked with the image capability.
• The model dropdown is empty with "请先到配置里添加渠道和模型": the administrator still needs to add a channel and its models.
• "生成图片缺少媒体存储键" or saving to assets fails: regenerate, and confirm the administrator configured default object storage.
• "生成连接已关闭，请重新生成": the streaming connection dropped (network blip or server restart); submit again.