SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsimage-gen

image-gen

Design & creation

View repository

Updated 10-15 20:29

About this skill

Image Gen

Single entry point for marketing image and AI video generation. All generation goes through the Higgsfield CLI — Use the model catalog (gpt_image_2, FLUX family, cinematic/video models) with one account. The default image model is gpt_image_2 (strong at reproducing text/wordmarks).

Hard rules

  1. Higgsfield is the only generation path. Direct API calls only in environments where Higgsfield is impossible (appendix), and even then, inform the user.
  2. Account check required before generation: higgsfield account status — Check the account email and credits before starting. Spending on the wrong account cannot be undone.
  3. Default model = gpt_image_2. Other models (FLUX, etc.) when specified by the user or needed for style characteristics — check the catalog with higgsfield model list --image. On failure, switch to another model — no silent switching — report and let the user decide.
  4. Variations by default generate multiple: at least 3 per use (different composition/tone/concept). Ad creative assumes A/B — 4+ (hook axis × visual axis). Generating 1 image violates this skill.
  5. Verify visually after generation: READ each result — text in the image character by character, finger/logo distortion, brand colors. For images where text is key, don't bake in the text — HTML/PIL overlayis the default. (Exception: YouTube thumbnails follow the thumbnail-maker rules — generating the copy together is the default, since gpt_image_2 is strong at text rendering.)
  6. Real people only based on user-provided photos (--image reference) — do not invent faces.
  7. On Windows, the CLI uses the full path: %APPDATA%\npm\higgsfield.cmd.

Setup (1회)

npm i -g @higgsfield/cli          # Node 필요
higgsfield auth login             # 브라우저 로그인
higgsfield account status         # VERIFY: 이메일 + 크레딧
higgsfield model list --image     # 카탈로그 확인 (gpt_image_2 포함)

Generation

# 텍스트 → 이미지
higgsfield generate create gpt_image_2 \
  --prompt "<장면 서술>" --aspect_ratio 16:9 --quality high --resolution 2k \
  --wait --json

# 이미지 입력 (실물 사진/디자인 레퍼런스 — 목업 합성, 인물 유지)
higgsfield generate create gpt_image_2 --image ./reference.png \
  --prompt "<이 이미지를 정확히 유지한 채 ...>" --aspect_ratio 9:16 \
  --quality high --resolution 2k --wait --json
  • Result parsing: Output is pretty-print JSON — result_url don't trust the key; recursively collect all image URLs from the entire stdout json.loads 후 recursively collect all image URLs, and URLs other than the host where the input reference was uploaded are the results. Output saved to a file gets a BOM (encoding='utf-8-sig'). Download with Python urllib.request.urlretrieve (avoids the Windows curl encoding trap).
  • Transient 502 → retry only once . Report repeated failures.
  • For composites where reference fidelity matters, specify in the prompt "reproduce EXACTLY as provided, same layout/text/colors, no redesign" — still, inform the user that fine print is an approximation (the source of truth is the original file).
  • AI video clips: check the model with higgsfield model list --videothen the same pattern — the editing/compositing pipeline is owned by ad-video and hyperframes.

Specs by use

Use Ratio Number of variations Notes
Newsletter hero 16:9 → resize to 1200 width 3 Text-free visual + title in the email HTML
Blog cover 16:9 3 Consider also for OG 1200×630 crop
Ad creative 1:1 / 16:9 per platform 4+ (A/B required) Different for hook axis and visual axis, run via paid-ads
Card news background 4:5 2–3 Specify in the prompt to leave the text area empty
Mockup compositing Match the scene 2+ --imageInput the actual design as, instruct to preserve EXACTLY

Workflow

  1. account status Check → check purpose/ratio/brand tone (DESIGN.md) → select model.
  2. Design variation axes (e.g., A=person / B=product close-up / C=abstract) → generate each.
  3. READ-verify all variations → regenerate failures (same model).
  4. Filename {용도}_{axis}_{n}.png — so A/B can be tracked.
  5. Present as a variation set + one line of axis explanation each + mention credit usage.

Appendix — Direct API (only for environments where Higgsfield is unavailable)

  • OpenAI gpt-image-2: POST /v1/images/generations (env OPENAI_API_KEY; billing_hard_limit_reached = report credit depletion). For editing, /v1/images/edits.
  • Google Nano Banana: Gemini generateContent + inline_data image input (env GEMINI_API_KEY). Both only after notifying the user. No other models/services allowed.