02 3d cgi
Model-aware Claude Agent Skills for AI image & video generation on Higgsfield — 15 creative skills with opt-in MCP generation. Reworked from beshuaxian/higgsfield-seedance2-jineng.
npx -y skills add pixelab-ch/higgsfield-skills --skill 02-3d-cgiAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Crafts 3D CGI and rendered video prompts for Higgsfield and routes to the best model (seedance_2_0 primary, wan2_7 fallback). Use when the user wants 3D rendered, CGI, Pixar-style, Unreal Engine, photorealistic 3D, computer- generated, or digitally rendered video. Triggers on: 3D animation, CGI, rendered, Blender, Unreal Engine, octane render, ray tracing, volumetric, subsurface scattering, physically based rendering, or any 3D/CG video request.
SKILL.md
6.1 KB, as published. Nobody here has run it
Higgsfield 3D CGI Skill
What this skill does
Crafts production-ready 3D CGI and rendered video prompts and routes them to the best Higgsfield model for computer-generated output. On request, it assembles the full parameter set, surfaces the credit cost, and — after explicit user confirmation — generates the asset via the opt-in generation flow.
The difference between a generic "make it look 3D" request and a production-ready CGI prompt lies in specificity: render engine terminology, material properties, lighting language, camera dynamics, and the architectural "hooks" that grab attention in the first two seconds.
Model routing
Primary and fallback video models
| seedance_2_0 (primary) | wan2_7 (fallback) | |
|---|---|---|
| Rationale | Broadest param set for CGI: genre tags, 21:9 aspect, full media-role support | Use only when primary is unavailable |
| Resolution | 480p, 720p, 1080p (def 720p) | 720p, 1080p |
| Aspect ratios | auto, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 | 16:9, 9:16, 1:1, 4:3, 3:4 — no 21:9 |
| Duration | 4–15 s (continuous range) | 2–15 s (continuous range) |
| Tunable params | resolution, mode {std,fast}, genre {auto,action,horror,comedy,noir,drama,epic} | None beyond resolution and aspect |
| Media roles | image, start_image, end_image, video, audio | start_image, end_image, audio only |
Routing rule: Use seedance_2_0 by default. Only fall back to wan2_7 when (1) the
primary model is unavailable, or (2) the user explicitly accepts wan2_7's constraints
after being told what they give up: no 21:9 aspect ratio, no genre or mode params,
and no image or video media roles. If the user has not been informed of these
trade-offs, stay on seedance_2_0 and let them decide — never switch silently.
MODEL-06 directive: If a parameter is rejected at generation time, call
higgsfield:models_explore with the target model name to re-verify the live schema.
Full parameter tables: references/model-specs.md.
Per-platform quick rule
Vertical 9:16 for TikTok / Reels / Shorts; 16:9 for YouTube and LinkedIn;
21:9 cinematic widescreen only on seedance_2_0. Full per-platform table:
references/model-specs.md.
Prompt-building workflow
-
Gather intent — Confirm with the user: CGI style (Pixar, photorealistic, low-poly, cel-shaded, isometric, etc.), subject, mood, target platform, aspect ratio, desired duration, and any I2V reference images.
-
Select model — Apply the routing table above. Choose
seedance_2_0by default; selectwan2_7only when a routing trigger applies. -
Build the prompt — Use the craft principles in the references below:
- Opening hook (2-second attention-capture technique) from references/hooks.md
- Render style, materials, lighting setup, camera movement from references/render-craft.md
- Worked example prompts by sub-genre from references/examples.md
-
Present for review — Show the assembled prompt and all parameters to the user for review and refinement before any generation call. The user must approve the prompt text, model, aspect ratio, duration, and (if used) genre/mode.
Opt-in generation
Generation costs Higgsfield credits and requires explicit user confirmation before any generate call is made. This skill never auto-generates.
Full step-by-step flow (confirmation gate, balance/cost surface,
generate → poll → display):
../../shared/generation-flow.md
This skill's primary model: seedance_2_0
Media upload: Conditional — only when the user supplies a start_image, end_image,
video, or audio reference. This skill is NOT a mandatory-I2V skill; T2V is the
default path. When the user provides a reference file, run the
higgsfield:media_upload → higgsfield:media_confirm atomic pair before generating.
Tool signatures: ../../shared/mcp-tools.md
Reference materials
| File | Contents |
|---|---|
| references/model-specs.md | Per-model parameter tables, aspect ratios, durations, routing rationale, and re-verify directive for seedance_2_0 and wan2_7 |
| references/render-craft.md | Render style encyclopedia (9+ styles), material science library (20+ surfaces), lighting setups (15+), camera movement guide (15+), particle and effects language |
| references/hooks.md | 2-second hook framework with 12+ hook techniques and prompt phrasing templates for CGI |
| references/examples.md | Master prompt template, 5 full production-quality CGI example prompts (Pixar, photorealistic product, isometric sci-fi, creature transformation, architectural flythrough), and common-mistakes guide |