Dreamifly Batch

Dreamifly 批量生图 / 生视频

把用户的提示词队列转成 Dreamifly 图片/视频生成任务，下载媒体到 images/，并用 results.jsonl / failed.jsonl 给 Agent 稳定汇报和重试。

Trigger Boundary

Use this skill for:

• 批量生图、图生图编辑、统一排队出图
• 文生视频、图生视频、多参考图生视频
• 校验 prompt 队列、估算积分、重试失败项、汇总生成结果

Do not use this skill for generic image prompt writing if the user does not want Dreamifly generation, local files, or queue execution.

Execution Protocol

Always run this sequence for real jobs:

1. Choose the model and write one task per line to prompts.txt, or use --prompts jobs.jsonl.
2. Run python3 dreamify.py --validate.
3. Run python3 dreamify.py --estimate.
4. Run python3 dreamify.py --dry-run.
5. Run python3 dreamify.py --check.
6. If any item is 5+ credits, video, or nano-banana-2, stop and ask for explicit user approval.
7. Run ./run.sh or python3 dreamify.py -n N.
8. Report with python3 dreamify.py --summary, then inspect results.jsonl and run.log if needed.

Cheap image models (Wai-*, Z-Image-Turbo, Qwen-Image-Edit) can run without extra approval after validation passes.

Queue Formats

Inline text:

a neon cat | model=Z-Image-Turbo | 16:9 | x2 | seed=123 | neg=blurry
edit this, add snow | model=Qwen-Image-Edit | img=ref.png
city timelapse | model=happyhorse-1.0 | secs=5 | res=720P

JSONL:

{"prompt":"a neon cat in rain","model":"Z-Image-Turbo","aspectRatio":"16:9","batch_size":1}
{"prompt":"edit this, add snow","model":"Qwen-Image-Edit","images":["ref.png"]}

Required Safety Rules

• Never print, commit, copy, or summarize config/cookie.txt; only say whether a cookie was loaded and its length if needed.
• Never run video, nano-banana-2, or any 5+ credit item without explicit approval.
• Do not blindly retry 401/402; ask the user to update the cookie or credits first.
• Video jobs are expensive and must not be auto-retried.

Progressive References

Read these only when needed:

• Model routing and parameters: references/model-selection.md
• Output files and reporting contract: references/output-contract.md
• Failure lessons and hard boundaries: references/gotchas.md
• Skill routing eval cases: references/evals.md
• JSONL examples: examples/prompts.jsonl

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: shaozheng0503
• Source: shaozheng0503/dreamifly-batch
• License: MIT

Install and usage instructions live in the source repository linked above.