multimodal-studio

name	Multimodal Studio
description	Unified image, video, and character generation across 30+ frontier models through a single connector. Model routing, visual prompt engineering, character consistency, async job handling, and brand-locked output for production creative work.
version	1.0.0
triggers	["generate image","generate video","create character","multimodal","higgsfield","image generation","video generation","b-roll","cinematic","thumbnail","hero image","character reference","visual asset"]

Multimodal Studio

One connector. Image + video + consistent characters. Thirty-plus models, vendor-agnostic, brand-locked.

Purpose

Turn a creative brief into finished visual assets — images, videos, and reusable characters — without leaving the agent loop. Multimodal Studio is ACOS's unified generation layer: it routes each request to the right model, engineers the prompt, holds brand and character consistency across a whole asset set, and handles the async generation lifecycle end to end.

This is the capability that closes the gap with agent-first platforms like Google Antigravity — except ACOS stays model-agnostic (no lock-in to one vendor's image/video stack) and brand-aware (every asset inherits Frank DNA + project brand tokens).

When to Use This Skill

• Generating a hero image, thumbnail, OG card, or social visual for a piece of content
• Producing short-form video, b-roll, or a cinematic clip from a still or a prompt
• Creating a reusable character that stays consistent across many images and videos
• Building a coordinated asset set (article header + 3 social cards + a 5s teaser) that must look like one campaign
• Any time a workflow says ~~image generation or ~~video generation

The Connector (Higgsfield MCP — default)

ACOS's default multimodal connector is Higgsfield MCP — one OAuth connection exposes 30+ image and video models plus character training. Per CONNECTORS.md, this skill is tool-agnostic: any MCP that fills the ~~image generation / ~~video generation categories works. Higgsfield is the default because a single server covers all three modalities with consistent characters.

Connect (Claude Code):

claude mcp add --transport http --scope user higgsfield https://mcp.higgsfield.ai/mcp
# then authenticate via OAuth in your Higgsfield account — no API keys to manage

Self-hosted alternative (stdio + API keys): see resources/model-matrix.md.

Tools exposed: generate_image, generate_video, create_character, list_characters, get_generation_status.

Verify before generating: if the higgsfield tools are not available, do NOT silently fall back to describing images in text. Tell the operator the connector is missing and give them the one-line claude mcp add command above.

The Workflow (BRIEF → ROUTE → CRAFT → GENERATE → ASSEMBLE)

1. BRIEF — lock intent before spending a generation

Capture (ask only for what's missing; infer the rest from project context):

• Goal & placement — where does this live? (blog hero, IG reel, YouTube thumbnail, OG card)
• Aspect ratio & resolution — derive from placement (see matrix); never ask if placement implies it
• Style — photoreal / 3D / illustration / minimalist / cinematic (inherit brand tokens if a brand skill is active)
• Character? — is there a recurring subject that must stay consistent across assets?
• Modality — still, video, or both (image → animate is the cheapest path to video)

2. ROUTE — pick the model deliberately (see resources/model-matrix.md)

Routing heuristics:

• Photoreal stills, people, products → Soul (4K, strong on faces/brand)
• Stylized / illustration / typography-in-image → Flux or Seedream
• Cinematic motion from a still → image-to-video (Kling / Hailuo / DoP)
• Text-to-video, complex scenes → Veo / Sora-class
• Recurring subject across the set → create_character once, then reference its ID in every call Always state which model you chose and why in one line before generating.

3. CRAFT — engineer the prompt

Structure every visual prompt as: Subject + Action + Setting + Composition + Lighting + Style + Technical.

• Lead with the subject and the single most important visual idea.
• Specify camera/lens/lighting for photoreal ("85mm, soft key light, shallow depth of field").
• For video, describe motion explicitly (camera move + subject motion + pacing); models default to static otherwise.
• Inject brand tokens (palette, mood, typography rules) when a brand skill is loaded.
• Negative space and aspect ratio are part of the prompt, not an afterthought.

4. GENERATE — handle the async lifecycle

• Generation is asynchronous. Submit, capture the job ID, then poll get_generation_status — do not assume instant results. Images: seconds. Video: tens of seconds to minutes.
• Submit independent assets in parallel (multiple generate_image calls in one turn), then poll. This is the multi-asset speed advantage.
• On failure or content-policy rejection: report it plainly, adjust the prompt, retry once — don't loop silently.
• Save returned assets to the project's canonical asset path; never leave them only as URLs.

5. ASSEMBLE — deliver a coherent set

• Verify the set reads as one campaign (same character ID, palette, lighting language).
• Produce required derivatives (e.g. crop hero → OG 1200×630, square 1080×1080).
• Log what was generated (model, prompt, job ID) so it's reproducible and auditable.

Character Consistency (the differentiator)

For any recurring subject (a brand mascot, a course instructor avatar, a series protagonist):

1. create_character once from a reference image or description → get a character ID.
2. Reference that ID in every subsequent generate_image / generate_video call.
3. list_characters to reuse across sessions and projects. This is how you get a character that looks identical across an entire content series — the thing single-shot image tools cannot do.

Quality Bar (Frank DNA)

• Show, don't tell — the asset is the proof; don't describe what you'll make, make it.
• Premium by default — highest sensible resolution, deliberate composition, no AI-slop tells (mangled hands, garbled text, plastic skin). Inspect output and regenerate if it has them.
• Brand-locked — if a brand skill is active, every asset obeys its palette/voice/tokens.
• Reproducible — log model + prompt + seed/job ID for every keeper.
• Cost-aware — image-to-video before text-to-video; lower res for drafts, 4K for finals.

Composes With

• frankx-brand / brand-guidelines → brand tokens for generation
• content-strategy / video-script → briefs that feed the studio
• suno-mastery → score the videos this skill produces
• infogenius command → research-grounded image prompts
• /studio and /generate-video commands → operator entry points

Reference

• resources/model-matrix.md — full model selection matrix, aspect-ratio table, self-hosted config, cost tiers