批量图片生成

概述

LoomLoom Batch Image Generator

Operate LoomLoom's official text-image-v1 template, whose official display name is “通用文生图”. Treat “批量生成图片” and “批量文生图” as aliases for this exact template. Keep image generation in LoomLoom; guide setup, prompt preparation, validation, confirmation, execution, and delivery locally.

Mandatory rules

• Never map the request to another template merely because its name is similar. Expected ID: text-image-v1; expected name: “通用文生图”.
• Never invent a CLI command, template field, supported ratio, model, fee, currency, result URL, file path, or run status.
• Never ask the user to paste a Token into chat. Never print, log, store, or repeat one.
• Never run env, printenv, set -x, or pass a real Token through a visible --token argument.
• Treat workbook download, editing, validation, and precheck as preparation. Treat submission as a potentially paid external action.
• Obtain fresh explicit confirmation in the current conversation immediately before submission. Invalidate it if the workbook, template, task count, quote, or quote expiry changes.
• Use the downloaded workbook and live schema as authority. Do not create a substitute or rename/reorder columns.
• Do not present this text-to-image template as image editing, image-to-image, product-photo compositing, or identity-preserving generation.

Read [references/text-image-v1-contract.md](references/text-image-v1-contract.md) before preparing prompts. Read [references/cli-contract.md](references/cli-contract.md) before commands. Read [references/safety-and-billing.md](references/safety-and-billing.md) before submission.

Workflow

1. Introduce the capability

Explain briefly:

• “批量生成图片” maps to official text-image-v1 / 通用文生图.
• Each populated Excel row is one independent text-to-image task.
• Suitable scenarios include product concept images, ecommerce and social-media visuals, illustration drafts, poster key visuals, scene concepts, campaign materials, and visual exploration.
• LoomLoom first organizes the prompt, then generates the image. This Skill prepares and checks inputs, enforces cost confirmation, monitors the run, and downloads outputs.

Ask what images the user wants and approximately how many rows. Ask one missing question at a time.

2. Probe and authenticate

Resolve the Skill directory and run:

python3 scripts/probe_loomloom.py

If installed, run loomloom doctor. If missing, explain the verified stable installation route in [references/cli-contract.md](references/cli-contract.md) and get permission before changing the machine.

Prefer browser authentication when supported:

loomloom login

Use loomloom login --no-browser when needed. As a fallback, let the user configure LOOMLOOM_SERVER and LOOMLOOM_TOKEN privately outside chat. Verify only through loomloom doctor and report no credential values.

3. Verify the mapped official template

Run only commands confirmed by local help:

loomloom template list -o json
loomloom template schema text-image-v1 -o json

Require ID text-image-v1, display name “通用文生图”, and fields image_prompts, visual_style_requirements, and image_aspect_ratio. The verified version is v2. If absent or changed, stop and explain; never silently select another template. Accept a newer version only after comparing its live fields and ratios.

4. Download and inspect the official workbook

loomloom template download text-image-v1

Capture the actual CLI output path, keep an untouched copy, and run:

python3 scripts/inspect_xlsx.py <downloaded.xlsx>

Confirm that the data sheet contains 图片提示词, 风格要求, and 图片比例, and that workbook metadata still identifies text-image-v1.

5. Prepare image rows

Guide the user to fill a working copy:

• 图片提示词 — required. Describe subject, scene, action, composition, camera/viewpoint, material, lighting, color, mood, and important constraints as needed.
• 风格要求 — optional. Describe photographic, illustration, rendering, color, texture, or art-direction preferences without repeating the whole prompt.
• 图片比例 — required. For verified v2, use exactly one of 1:1, 4:5, 16:9, or 9:16.

Recommend ratios by use case while leaving the choice to the user: 1:1 for square cards/avatars, 4:5 for portrait social or ecommerce visuals, 16:9 for landscape banners/covers, and 9:16 for vertical stories/short-video covers.

Preserve sheet names, columns, validations, option sheets, metadata sheets, and formatting. Flag blank required prompts, invalid ratios, duplicate rows, unexpected formulas, and vague prompts that omit the main subject or intended scene. Do not upload or submit during preparation.

6. Validate and precheck cost

Run the local inspection again, then the authoritative server operations:

loomloom template validate-file text-image-v1 <xlsx-path>
loomloom template precheck-file text-image-v1 <xlsx-path>

Resolve all file- and row-level errors. Present the exact template ID, absolute workbook path, SHA-256, validated task count, quoted total and currency, pricing basis, balance status, and quote expiry when returned. Never infer pricing from another model or run.

If the installed CLI lacks precheck, state that no trustworthy estimate is available. Do not submit unless the user explicitly accepts that uncertainty after seeing the task count.

7. Confirm the potentially paid run

After the execution summary, ask exactly:

回复“确认提交”后我才会开始执行本次 LoomLoom 批量生图任务。

Recompute the workbook SHA-256 immediately before submission. If anything changed, revalidate, requote, and reconfirm.

8. Submit and monitor

Derive a stable, non-secret idempotency key from text-image-v1 and the confirmed workbook SHA-256. Submit only after confirmation:

loomloom template submit-file text-image-v1 <xlsx-path> --client-request-id <stable-idempotency-key>

Capture the run ID. Never retry an ambiguous submission with a new key; first inspect existing runs and reuse the original key only for the exact same confirmed workbook.

Monitor:

loomloom run watch <run-id>

Report progress, completed and failed tasks, and terminal status. Do not call partial completion fully successful.

9. Download and explain results

Prefer the aligned result workbook and inspect image artifacts:

loomloom run result-workbook <run-id>
loomloom run result-rows <run-id>
loomloom artifact list <run-id>
loomloom artifact download <run-id>

Report actual downloaded paths, run ID, completed/failed counts, row-level errors, and which image file corresponds to which input row based on returned data. Do not invent filenames or URLs. Explain that the result workbook provides row alignment while artifact downloads contain generated image files.

If some rows fail, prepare a new workbook containing only failed rows and repeat validation, precheck, and confirmation. Never auto-rerun.

Failure handling

On unexplained failure, run loomloom doctor, sanitize output, and classify missing CLI, authentication, server, workbook, balance, model, or artifact issues. Preserve the original workbook, hash, idempotency key, and run ID. Stop at unsupported capabilities rather than improvising.