메뉴
Generate consistent photo sets using a model card for face-locking. Two modes — Grid Mode (4-photo grid) and Chain Mode (sequential, unlimited). Use when the user asks for photo sets, beauty shots, outfit photos, or套图.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/fang-lin/skills --skill consistent-portrait-set명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
SOC 직업 분류 기준
SKILL.md 표시 중
| name | consistent-portrait-set |
| description | Generate consistent photo sets using a model card for face-locking. Two modes — Grid Mode (4-photo grid) and Chain Mode (sequential, unlimited). Use when the user asks for photo sets, beauty shots, outfit photos, or套图. |
| version | 0.8.0 |
| license | MIT |
| compatibility | Requires GEMINI_API_KEY, internet access, Python packages google-genai and Pillow. |
| platforms | ["macos","linux","windows"] |
| required_environment_variables | [{"name":"GEMINI_API_KEY","prompt":"Gemini API key for image generation","help":"https://aistudio.google.com/app/apikey","required_for":"all image generation and upscaling operations"}] |
| metadata | {"author":"fang-lin","repo":"https://github.com/fang-lin/consistent-portrait-set","hermes":{"tags":["photo","portrait","image-generation","face-lock","gemini","套图","consistency"],"category":"creative","requires_toolsets":["code_execution","skills","web"],"config":["[Truncated]","[Truncated]","[Truncated]"]}} |
Generate a set of 4 consistent photos from a single model card (face reference), with customizable outfits, environments, and accessories.
[!IMPORTANT]
- ALL steps MUST be executed through
generate_suite.pyscript. Do NOT write your own code to call Gemini API, crop images, or upscale. The script handles directory structure, logging, and file management.- Script location: Find it at
the skill scripts directory (find via skill_view)or~/.hermes/skills/consistent-portrait-set/scripts/generate_suite.py- Send images via
MEDIA:prefix only. Never use browser_navigate or any other method to show images.- Three approval gates: creative brief (Step 1), prompt (Step 4), and grid image (Step 6). Do NOT skip any.
- All intermediate and output files are managed by the script. Do not save files to /tmp or any other location manually.
- Don't patch skill files yourself. If you find a bug or missing feature in this skill, tell the user and suggest running
hermes skills update consistent-portrait-set. The skill is maintained upstream — local edits will be overwritten on the next update.
Before anything else, understand what the user wants and produce a creative brief.
Two modes — choose based on the user's input:
Mode A: the user gives a vague request (e.g. "给我出一套图", "来点新的") → You take the initiative. Use web_search to check current trending topics, seasonal themes, fashion trends. Draft a complete creative brief and present it for approval.
Mode B: the user gives a specific direction (e.g. "我想要一套职业装的", "穿红裙子在海边") → You ask clarifying questions to fill in the gaps, then confirm the brief.
The brief format:
📸 套图策划
名称：[套图名称]
主题：[主题/风格关键词]
热点关联：[当下相关的热度话题、节日、流行趋势，没有就写"无"]
场景：[拍摄场景描述]
服装：[服装方向描述]
调性：[清新 / 性感 / 职业 / 街拍 / 文艺 / 暗黑 / ...]
参考说明：[简述创意思路，1-2 句话]
STOP and WAIT for the user to approve, modify, or reject the brief.
Required:
the assets directory for available references, or ask the user to send one.Optional (ask the user if he wants to provide):
Build a detailed English prompt for Gemini image generation. The prompt MUST include:
Prompt template:
A strictly equal 2x2 grid of 4 portrait photos of [person description]. The grid must have exactly 4 cells of equal size, arranged in 2 rows and 2 columns, with no overlapping or irregular layout. Each cell contains one full portrait photo.
Top-left: [pose/angle 1], [outfit detail], [scene detail]
Top-right: [pose/angle 2], [outfit detail], [scene detail]
Bottom-left: [pose/angle 3], [outfit detail], [scene detail]
Bottom-right: [pose/angle 4], [outfit detail], [scene detail]
Style: [lighting], [mood], professional photography, high resolution, sharp details.
[!IMPORTANT] You MUST show the complete prompt to the user BEFORE generating. Do NOT call the API without the user's approval.
Send the composed prompt to the user in a clear format:
我准备用以下 prompt 生成四宫格，你看看：
[你组装的完整英文 prompt]
参考图：
- 模特卡：[路径]
- 环境：[路径 or 无]
- 服装：[路径 or 无]
- 饰物：[路径 or 无]
分辨率：4K / AR：9:16
可以吗？
STOP and WAIT for the user to approve, modify, or reject.
MUST use the script. Do NOT call Gemini API directly.
Default resolution is 4K. The script will output the grid to tmp/<session-id>/grid.png.
from hermes_tools import terminal
SCRIPT = "the skill scripts directory (find via skill_view)"
result = terminal(
f'python {SCRIPT} '
'--step generate '
'--model-card /path/to/model-card.jpg '
'--prompt "your composed prompt here" '
'--session-id 20260426_120000'
# Optional reference images — include only if the user provided them:
# ' --env-ref /path/to/env-ref.jpg'
# ' --outfit-ref /path/to/outfit-ref.jpg'
# ' --accessory-ref /path/to/acc-ref.jpg'
# Optional resolution override (default 4K):
# ' --resolution 2K'
)
print(result["output"])
[!IMPORTANT] This step is MANDATORY. You MUST send the grid to the user IMMEDIATELY after generation. Do NOT skip this step. Do NOT proceed to crop or any other step without the user's explicit approval.
Send the grid image to the user using MEDIA: prefix (this is Hermes's native way to send files in Telegram — just include the path with MEDIA: prefix in your response text):
MEDIA:<path from script output>/grid.png
Ask: "这组四宫格你满意吗？满意我就裁切出图了。"
STOP and WAIT for the user to approve before proceeding.
If the user says no:
MUST use the script.
result = terminal(
f'python {SCRIPT} '
'--step crop '
'--session-id <same session id>'
)
print(result["output"])
If the 4K grid was used (default), cropped images are already ~2K each. Move them directly to output:
result = terminal(
f'python {SCRIPT} '
'--step finalize '
'--session-id <same session id>'
)
print(result["output"])
Only use upscale if the user explicitly asks for higher resolution or better detail. Each image can have different settings.
Upscale prompt guidelines:
Reference images during upscale are OPTIONAL and per-image. Ask the user if he wants to add any.
# Upscale all 4 images to 4K
result = terminal(
f'python {SCRIPT} '
'--step upscale '
'--session-id <same session id> '
'--resolution 4K '
'--upscale-prompt "your upscale prompt"'
)
print(result["output"])
# Or upscale a specific image with custom ref and resolution
result = terminal(
f'python {SCRIPT} '
'--step upscale '
'--session-id <same session id> '
'--resolution 4K '
'--upscale-prompt "your upscale prompt" '
'--upscale-ref /path/to/ref.jpg '
'--image-index 2'
)
print(result["output"])
Send all 4 final images to the user using MEDIA: prefix. Use the output directory path from the finalize/upscale step:
MEDIA:<output-dir>/final-1.png
MEDIA:<output-dir>/final-2.png
MEDIA:<output-dir>/final-3.png
MEDIA:<output-dir>/final-4.png
An alternative to Grid Mode. Generates images one at a time, each with its own approval. No limit on number of images. Best for higher consistency and fine control.
Let the user choose, or suggest based on context.
Step 1–2: Same as Grid Mode (Creative Brief + Gather Inputs)
Step 3: Generate Image 1
from hermes_tools import terminal
SCRIPT = "the skill scripts directory (find via skill_view)"
result = terminal(
f'python {SCRIPT} '
'--step chain '
'--model-card /path/to/model-card.jpg '
'--prompt "your prompt for image 1" '
'--session-id spring_casual_001'
# Optional: --resolution 4K (default 2K for chain mode)
)
print(result["output"])
Send to the user: MEDIA:<session-dir>/chain-1.png
Ask: "第 1 张你满意吗？继续下一张？"
STOP and WAIT for approval.
Step 4+: Generate subsequent images
the agent decides which reference images to pass based on content:
result = terminal(
f'python {SCRIPT} '
'--step chain '
'--model-card /path/to/model-card.jpg '
'--ref-images "<session-dir>/chain-1.png,<session-dir>/chain-3.png" '
'--prompt "your prompt for next image" '
'--session-id spring_casual_001'
)
print(result["output"])
Send to the user, wait for approval. Repeat.
Final step: Finalize
When the user says enough:
result = terminal(
f'python {SCRIPT} '
'--step chain-finalize '
'--session-id spring_casual_001'
)
print(result["output"])
Deliver all finals with MEDIA: prefix.
Choose references based on what you want to preserve:
| Goal | Which image to pick as reference |
|---|---|
| Lock face | One with clear, front-facing portrait |
| Lock outfit | One with full-body shot showing complete outfit |
| Lock scene | One with the desired background/environment |
| Lock pose style | One with similar body language |
| Maximum diversity | Only pass model card, no chain images |
[!IMPORTANT] Do NOT mechanically pass "the previous image". Choose references intentionally based on what needs to stay consistent. If a chain image is a close-up, it has no outfit info — don't use it to lock outfit.
All files are managed by the script under the workspace directory (auto-created by script):
workspace/consistent-portrait-set/
├── outputs/ # Final deliverables (by date)
│ └── YYYY-MM-DD/
│ └── <session-id>/ # e.g. colmar_001, beach_sunset_002
│ ├── final-1.png
│ ├── final-2.png
│ ├── final-3.png
│ └── final-4.png
├── tmp/ # Intermediate files (can be cleaned)
│ └── <session-id>/
│ ├── grid.png # Grid Mode output
│ ├── crop-1~4.png # Grid Mode crops
│ ├── chain-1.png # Chain Mode images
│ ├── chain-2.png
│ └── chain-N.png
└── logs/
└── YYYY-MM-DD.log # All API calls logged
Grid Mode:
| Grid Resolution | Cost | Crop Size (each) | Need Upscale? |
|---|---|---|---|
| 4K (default) | $0.151 | ~2K | No — good for social media |
| 2K | $0.101 | ~1K | Maybe — depends on use case |
| 1K | $0.067 | ~500px | Yes — too small without upscale |
Chain Mode:
| Resolution | Cost per image | Notes |
|---|---|---|
| 2K (default) | $0.101 | Good for social media |
| 4K | $0.151 | When higher quality needed |
references/prompt-guide.md — 详细的提示词模板、调性参考、服装/场景/光线描述技巧、手脚质量控制、超分指南。每次执行 consistent-portrait-set 前建议先加载：skill_view("consistent-portrait-set", file_path="references/prompt-guide.md")TypeError: 'NoneType' object is not iterable on response.candidates[0].content.parts, the prompt was silently blocked by Gemini's content moderation. See the rewriting table below.When the user asks for revealing or sexy outfits (大尺度, 性感, lingerie), Gemini's safety filters will block explicit descriptions. Rewrite using euphemistic/artistic vocabulary:
| Blocked (too direct) | Allowed (euphemistic) |
|---|---|
| lingerie / bodysuit | figure-hugging mini dress, satin slip dress |
| deep plunging neckline | flattering low neckline |
| sheer / transparent / see-through | translucent |
| cleavage | low neckline, revealing neckline |
| barely covering | short hemline, mini length |
| skimpy | form-fitting, body-hugging |
| naked / topless | bare shoulders, off-shoulder |
| lace lingerie set | black satin robe loosely draped |
| straps falling off | robe hanging loosely off one shoulder |
Principles:
tmp/<session-id>/grid.png exists before sending to user.outputs/YYYY-MM-DD/<session-id>/.MEDIA: prefix