convez.ai
SkillAutomaçãoAvançado
wifiwave

Higgsfield Content Pipeline

wifiwave/higgsfield-content-pipeline

Automação de produção de vídeo com geração de stills e animações via IA

Ver no GitHub
InicialPluginScore · de 0 a 100

Studio de produção de vídeo automatizado que integra Claude Code com Higgsfield para gerar reels, anúncios de e-commerce e vídeos 360°. Inclui 19 skills prontas, playbook de prompting, manual operacional e prompts de produção testados em campo.

Instalação

$npx convez-ai wifiwave/higgsfield-content-pipeline

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
34
Forks
2
Acompanhando
2
Saúde
42%
Último commit
há 2 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Gerar reels cinematográficos de marca pessoal com treino de identidade facial
  • Criar anúncios de e-commerce e imagens de produtos automaticamente
  • Produzir cards de marketplace e vídeos 360° de produtos
  • Montagem automática de sequências de vídeo com aprovação de keyframes

Veredito

Pipeline robusto e pronto para produção, com 19 skills testadas em reels reais. Exige Higgsfield Pro e Claude Code, mas elimina a repetição manual de geração e montagem de vídeos. Recomendado para criadores e negócios de e-commerce que já investem em IA generativa.

README

A tradução para o português deste README ainda está na fila.

Higgsfield Content Pipeline

Turn Claude Code + Higgsfield into an automated video production studio.

Train an identity model on your face once. After that, Claude writes the prompts, fires the renders, checks the keyframes, and assembles the final cut. You approve stills and pick takes. The same pipeline that makes cinematic personal-brand reels also makes e-commerce product ads, listing images, and 360 product videos.

Everything in this repo came from real production runs (shipped, sponsored Instagram Reels), not theory. The prompt frameworks, the IP-filter workarounds, and the credit costs below are all battle-tested.

What's inside

PathWhat it is
skills/4 core skills that drive the official higgsfield CLI: generate, soul-id, product-photoshoot, marketplace-cards
skills/seedance/15 Seedance 2.0 style skills (cinematic, e-commerce ad, product 360, fashion, food, real estate, and more)
playbook/The Seedance prompting playbook, photo reference guide, and production checklist
CLAUDE.mdThe agent operating manual. Claude Code reads this automatically and runs the pipeline for you
examples/prompts/Real production prompts (still + motion pairs) from a shipped 5-scene cinematic reel
install.shSymlinks all 19 skills into ~/.claude/skills

Requirements

  1. Claude Code (or any agent that reads SKILL.md skills)
  2. A Higgsfield account. The Plus plan ($29/mo, ~1,210 credits) is enough to ship a 5-scene reel with retries
  3. The higgsfield CLI (install command below)
  4. ffmpeg for assembling clips (brew install ffmpeg on macOS)

Quick start

git clone https://github.com/wifiwave/higgsfield-content-pipeline.git
cd higgsfield-content-pipeline
./install.sh                # symlinks all skills into ~/.claude/skills

# install + authenticate the Higgsfield CLI
curl -fsSL https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh | sh
higgsfield auth login

# open Claude Code in this repo and just ask
claude
# "train my Soul"  →  "make me a 3-scene cinematic reel"  →  done

The pipeline (personal brand / faces)

Soul-2 training  →  stills  →  Seedance 2.0 I2V  →  ffmpeg concat
   (once)          (~15 cr)    (~150 cr / 10s)       (free)
  1. Train Soul-2 once. 8 to 15 front-facing photos, no side profiles, no filters. Takes ~10 minutes, returns a UUID you reuse forever.
    higgsfield soul-id create --name me --soul-2 --image p1.jpg --image p2.jpg ...
    
  2. Generate a still per scene. ~15 credits, ~1 minute each.
    higgsfield generate create text2image_soul_v2 --prompt "$(cat still.txt)" \
      --soul-id <UUID> --quality 2k --aspect_ratio 9:16 --wait
    
  3. Animate it with Seedance 2.0. ~150 credits per 10s clip, 5 to 10 minutes.
    higgsfield generate create seedance_2_0 --prompt "$(cat motion.txt)" \
      --start-image still.png --duration 10 --aspect_ratio 9:16 --wait
    
  4. Concat the final picks.
    ffmpeg -f concat -safe 0 -i list.txt -c copy reel.mp4
    

You don't run these by hand. The skills + CLAUDE.md mean Claude runs the loop and brings you keyframes for approval.

For e-commerce

This is where the pipeline prints money for product businesses. No Soul training needed, just product photos or a product URL.

You wantUse
Studio / lifestyle product photoshiggsfield-product-photoshoot
Marketplace listing set (main image, secondary images, A+ content modules)higgsfield-marketplace-cards
Product demo / hero video adseedance-ecommerce-ad
360° rotating product shotseedance-product-360
Food & beverage hero footage, pour shotsseedance-food-beverage
Fashion campaign, model walksseedance-fashion-lookbook
3 to 7 second scroll-stopper hooksseedance-social-hook
30 to 60 second brand narrativeseedance-brand-story
Branded ads with avatars / UGC stylehiggsfield-generate (Marketing Studio)

Example asks once installed: "make a 360 video of this water bottle", "build the full marketplace listing image set for this product photo", "make a 10 second UGC-style ad for this skincare jar".

Hard-won lessons (read before burning credits)

  • The IP filter is the biggest landmine. Seedance blocks anything that pattern-matches copyrighted franchises or real celebrities. Strip franchise names and distinctive terms from prompts; describe the vibe generically ("ancient Greek warrior in a throne room with red cape" works, "Spartan from 300" gets blocked). If a Soul-generated still itself keeps failing, switch from --start-image to --image reference mode.
  • Soul training overfits to your most distinctive photo. Use 8 to 10 clean front-facing shots. One straight-on, no-filter photo is the anchor. Details in playbook/PHOTO-REFERENCE-GUIDE.md.
  • Always --soul-2. Soul-Cinematic is a different training type and will give you someone else's face.
  • Fix the big-head problem. Soul stills bias toward face dominance. For full-body shots, prompt a low-angle hero framing with the face at 12 to 15% of frame.
  • End every video prompt with "NO MUSIC. Native diegetic sound only." Otherwise each clip generates its own soundtrack and your concat has 5 clashing scores. Lay one music bed in post.
  • Multi-shot beats read as cinema, static portraits read as AI. Break every 10s prompt into 4 to 6 timestamped micro-beats (the examples show this).
  • --wait times out at 10 minutes but the job keeps running. Check higgsfield generate list before re-firing, or you'll pay twice.

The 7-part prompt framework

Every still prompt in examples/prompts/ follows this structure:

  1. Technical setup (resolution, aspect ratio, fps)
  2. Subject identity lock ("the character in my Soul profile, as [role]")
  3. Wardrobe and physical details
  4. Setting (what's behind and around)
  5. Lighting and color (color temperatures, real film references)
  6. Camera and composition (lens, aperture, angle)
  7. Motion sequence with timestamped beats (video prompts only)

Plus a closing identity + safety lock (restate the face reference, "NO blood", "NO trademarked brands", "NO MUSIC"). Full reference in playbook/PROMPTING-PLAYBOOK.md.

Real cost numbers

ItemCreditsWall time
Soul-2 training (once)~50~10 min
Soul still (2K)~15~1 min
Seedance 10s clip~1505 to 10 min
5-scene reel with retries~900 to 1,100one evening

The $29/mo Plus plan (~1,210 credits) covers roughly one full reel per month. Heavier production wants a bigger plan.

Credits and licenses

  • The 15 skills in skills/seedance/ are vendored from beshuaxian/higgsfield-seedance2-jineng (MIT). See skills/seedance/ATTRIBUTION.md.
  • The 4 higgsfield-* skills wrap the official Higgsfield CLI and are distributed as part of the Higgsfield agent tooling (v0.3.0).
  • Everything else (playbook, pipeline manual, examples, installer): MIT, see LICENSE.

Built by Agentic Solution. If this repo gets you signed up, use the referral link so the attribution lands.

Ver o original no GitHub

Bom para

Relacionados