by mojomast
Agent-agnostic skills that turn one prompt into a video with open-weights ComfyUI workflows (LTX, FLUX Klein, Qwen, Wan) - video-director orchestrator + 35 skills. #OpenSourceHiggsfield
Available after the catalog security scan passes.
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM lists this project for discovery and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
The deep catalog scan for this skill is still queued. Run an instant dependency check now instead.
# Add to your Claude Code skills
git clone https://github.com/mojomast/open-higgsfield-skillsGuides for using ai agents skills like open-higgsfield-skills.
See how open-higgsfield-skills compares with popular alternatives.
▶ Watch the demo (MP4, 1:26) · captions (SRT) · details
One prompt in, a finished short video out, using only open-weights models on your own ComfyUI.
open-higgsfield-skills is an agent-agnostic pack of 37 Agent Skills that teaches any coding agent (Claude Code, Codex, Cursor, OpenCode, Hermes, Pi or your own) to work as a film crew. The video-director skill takes a one-line idea such as "a lighthouse keeper finds a message in a bottle" and breaks it into a story bible, a shot list, keyframes, video shots, voice, music, captions and a final cut. It sends every step to ComfyUI workflows built on open weights: FLUX Klein, LTX-2.x, Qwen-Image-Edit, Wan, SeedVR2, ACE-Step, Chatterbox, Whisper and more.
Status: early and community-built. The workflows come from #OpenSourceHiggsfield hackathon projects and were checked statically (graph structure, node and weight requirements, routing and dry runs). Most have not been run end to end on a GPU. The routes in the demo were run on a Google Colab GPU. See Status and limitations.
Every shot in the demo video was rendered by this pack from one manifest. video-director validated and packaged it, and google-colab-runner ran it as a single notebook on a Google Colab RTX PRO 6000 (96 GB, standard tier). All 14 jobs finished in about 8.5 compute units, including setup and the roughly 99 GB model download. The narration was also generated by the pack.
| shot | what you see | skills and routes |
|---|---|---|
| s01 | astronaut on a red desert ridge | flux3-klein-stills (still.klein) → ltx25-image-to-video (video.i2v, two-pass + latent upscale) |
| s02 | the astronaut walks on | video-director last-frame chaining (chain_from: s01) → video.i2v |
| s03 | fox in a yellow raincoat; rain turns to snow; cut-out | still.klein → video.i2v, image-edit (fix.edit, Qwen-Image-Edit 2511), background-matte (fix.matte, BiRefNet) |
| s04 | vintage film camera, arc move | still.klein → camera-control (camera.fun, Wan Fun Camera, MOVES: "arc-left-45") |
| s05 | lighthouse in a storm | still.klein → video.i2v |
| s06 | robot painting in a studio | still.klein → video.i2v |
| VO | the narration | voiceover (audio.voice, Chatterbox Multilingual) |
The final cut, captions and title cards were made with ffmpeg rather than film-assembly, and still-qa, shot-qc, continuity-check and music were not exercised in this run. LTX-2.5 weights are under the Lightricks LTX-2 Community License, which is not OSI-approved.
Hosted "AI director" products string together closed models behind a paywall. This pack does the same job in the open:
video-director plans, renders, judges, fixes and assembles, and records every prompt, seed, model and cost in one resumable JSON manifest.SKILL.md plus stdlib-only Python scripts and ComfyUI API-format workflows, so any agent that reads Agent Skills can use them.flowchart TD
A[One-line prompt] --> B[Intake: smart defaults, at most 3 questions]
B --> C[comfyui-preflight + tier pick<br/>draft 16 GB / standard 24 GB / max 40 GB]
C --> D[story-bible]
D --> E[shot-planning]
E -->|approval gate| F[Stills: flux3-klein-stills]
F --> G{still-qa judge}
G -->|fail: identity / framing / light| H[image-edit / relighting / re-prompt / seed+1]
H --> G
G -->|pass| I[Video per shot]
I --> I1[ltx25-image-to-video]
I --> I2[lip-sync]
I --> I3[restyle / one-take-vfx]
I --> I4[video-inpaint-remove / character-swap-h3]
I --> I5[control-to-video / camera-control<br/>audio-driven-motion / ltx-iclora-utils / effect-recipe-read]
E -.-> PV[previs-control-passes] -.-> I5
E -.->|recurring identity, human dataset gate| LT[lora-training]
I1 & I2 & I3 & I4 & I5 --> SQ[shot-qc] --> J{continuity-check}
J -->|break| K[regenerate or shot-transitions]
K --> J
J -->|ok| L[Audio: voiceover + music-sfx]
L --> M[Post: camera-lens-lut, video-upscale, frame-interpolation, background-matte]
M --> N[film-assembly + subtitles-captions]
N --> O[reframe-aspect for 9:16 / 1:1 deliverables]
O --> P[Critique + REPORT.md]
subgraph transport[Every render]
Q[comfyui-connect] --> R[(ComfyUI server)]
Q -.-> MG[multi-gpu-dispatch: several servers]
end
Routing is data-driven: 70 routes today. skills/video-director/references/routing.md maps each need (for example "speaking on camera" or "remove object") to a skill, a workflow, its input bindings, a fallback route and the tiers it suits. To add a skill, you add one row.
Licence = the licence of the code and workflows in that folder. Model weights are never bundled and carry their own licences. They are mostly unverified here, so check the model card before commercial use. VRAM figures are estimates from upstream projects unless noted, not measurements by this repo.
| Skill | What it does | Models | VRAM | Licence |
|---|---|---|---|---|
audio-driven-motion |
Make video whose motion is driven by an audio track (a talking or singing avatar with body motion, or music-reactive motion) using open-weights LTX-2 audio-to-video, LTX-2.3 image+audio-to-video, Wan 2.2 S2V, InfiniteTalk or HuMo in ComfyUI | LTX-2 a2v, LTX-2.3, Wan 2.2 S2V, InfiniteTalk, HuMo | 24 GB+ (est.) | Apache-2.0, MIT · weights: LTX-2 Community License (not OSI); Wan S2V / InfiniteTalk: Apache-2.0 (unverified); HuMo: UNVERIFIED; MelBand RoFormer: UNVERIFIED |
background-matte |
Remove or replace backgrounds and pull mattes with open-weights BiRefNet in ComfyUI (stills to transparent or solid colour, video background removal), plus PlateLock plate compositing (difference matte, temporal stabilise, fidelity report) to put a generated element back onto a real plate | BiRefNet | ~2-4 GB (est.) | MIT · weights: BiRefNet: MIT (unverified) |
camera-control |
Animate a still with an explicit, repeatable camera move (named trajectories such as orbit, dolly or crane, or drawn ATI point tracks) using open-weights Wan Fun Camera and Wan ATI in ComfyUI | Wan 2.1/2.2 Fun Camera, Wan ATI | ~8-12 GB (1.3B); 24 GB+ (14B) (est.) | Apache-2.0, MIT · weights: Wan 2.1/2.2 Fun Camera, ATI: Apache-2.0 (unverified) |
camera-lens-lut |
Apply film-style camera moves (handheld, push, drift), lens effects, a 3D .cube grade LUT and letterboxing to a clip in ComfyUI with the bundled Backlot node pack | none (Backlot post nodes, .cube LUTs) | light (unverified) | MIT · weights: no model weights in the Backlot nodes; loaders unverified |
character-swap-h3 |
Swap a person, face or object in an existing clip for a reference image using open-weights MiniMax H3 (local weights) with D3voz's one-click swap nodes, or an open LTX-2.3 motion-swap fallback | MiniMax H3 (local), Qwen3-VL; LTX-2.3 fallback | 32 GB+ (est.) | MIT · weights: MiniMax H3: UNVERIFIED; Qwen3-VL: Apache-2.0 (unverified); LTX-2 Community License (fallback, not OSI) |
comfyui-connect |
Submit ComfyUI API-format workflows to any ComfyUI server (local, a Google Colab tunnel, or a hosted instance), poll for completion, download outputs, upload inputs, and cancel jobs | - | - | MIT · weights: none |
comfyui-preflight |
Check, before submitting, that a ComfyUI server has every node type and model weight file a workflow needs. Reads /object_info only; submits nothing | - | - | MIT · weights: none |
continuity-check |
Sweep adjacent shots and takes for continuity breaks (cast, wardrobe, props, grade, lighting, screen direction, timing) at plan level and take level, and report findings with proposed fixes - never edits | open VLM via still-qa | VLM-dependent | MIT · weights: uses still-qa VLM |
control-to-video |
Generate video that follows a control signal (depth, OpenPose skeleton or canny edges extracted from a reference or previs clip) with open-weights LTX-2 and Wan 2.2 Fun Control / Wan VACE in ComfyUI | LTX-2, Wan 2.2 Fun Control, Wan VACE, DWPose | 24 GB+ (est.) | Apache-2.0, MIT · weights: LTX-2 Community License (not OSI); Wan 2.1/2.2: Apache-2.0 (unverified); DWPose: Apache-2.0; SDPose: UNVERIFIED; Depth Anything >Small: CC-BY-NC (not bundled) |
effect-recipe-read |
Reverse-engineer a video effect into a reusable JSON recipe: a local Qwen3-VL 8B reads START/MIDDLE/END frames of |
open-higgsfield-skills is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by mojomast. Agent-agnostic skills that turn one prompt into a video with open-weights ComfyUI workflows (LTX, FLUX Klein, Qwen, Wan) - video-director orchestrator + 35 skills. #OpenSourceHiggsfield. It has 51 GitHub stars.
open-higgsfield-skills's catalog security scan is still queued. You can run an instant dependency and prompt-injection check now with the "Scan for vulnerabilities" button above.
Clone the repository with "git clone https://github.com/mojomast/open-higgsfield-skills" and add it to your Claude Code skills directory (see the Installation section above).
open-higgsfield-skills is primarily written in Python. It is open-source under mojomast on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh open-higgsfield-skills against similar tools.
No comments yet. Be the first to share your thoughts!