---
name: venice-studio-production
description: End-to-end AI video production workflow combining Hermes Agent (Telegram), Venice API (bulk generation), and Venice Studio Web (final assembly). Covers character consistency, multi-shot generation, audio scoring, and the hybrid API+UI pipeline.
version: 1.0.0
author: Hermes Agent
license: MIT
metadata:
  hermes:
    tags: [venice, video-production, ai-video, telegram, creative-workflow, seedance]
    related_skills: [venice-video, venice-image-generate, venice-audio-music, venice-chat, venice-models]
---

# Venice Studio Production Workflow

Produce AI videos from concept to export by combining three layers:

1. **Telegram + Hermes Agent** — Natural-language direction and bulk asset generation
2. **Venice API** — Programmatic image/video/audio generation with character consistency
3. **Venice Studio Web** (`venice.ai/studio`) — Timeline editing, transitions, audio sync, Seedance 2.0 cinematic shots

## Trigger

Load this skill when the user wants to:
- Create videos from Telegram prompts
- Generate character-consistent multi-shot scenes
- Produce AI micro-dramas, trailers, or episodic content
- Combine API-driven bulk generation with manual Studio editing
- Understand where Hermes ends and Venice Studio begins

## Architecture

```
Telegram Prompt → Hermes Agent → Venice API → Raw Assets
                                      ↓
                              Venice Studio Web
                         (Movie Editor + Seedance 2.0)
                                      ↓
                                Final Export
```

| Layer | Role | Key Tools |
|---|---|---|
| **Hermes / Telegram** | Concept, treatment, iteration, QA | `venice-chat`, `venice-image-generate`, `venice-video`, `venice-audio-music` |
| **Venice API** | Bulk generation, character consistency, parallel queuing | `elements[]`, `reference_image_urls[]`, `/video/queue` |
| **Venice Studio Web** | Timeline assembly, transitions, audio sync, cinematic polish | Movie Editor, Seedance 2.0 R2V, Library |

## Phase 1: Concept & Treatment (Telegram → Hermes)

### Start with a prompt

User sends something like:

> "Make a 60-second sci-fi trailer about an AI that achieves consciousness"

Hermes agent loads `venice-chat` and generates:
- Full treatment (concept, characters, setting, genre, tone)
- Shot list with descriptions, camera angles, duration
- Character reference prompts (for image generation)
- Scene reference prompts (environments, backdrops)
- Music score timing cues

### Treatment structure

```yaml
series:
  title: "The Oracle"
  genre: sci-fi thriller
  tone: mysterious, cerebral
  aesthetic: neon-noir, volumetric fog
  aspect_ratio: "16:9"
  target_duration: "60s"

characters:
  - name: Srikhav
    role: protagonist
    description: "Former AI researcher, haunted"
    voice: "Gravelly, exhausted"

shots:
  - id: 1
    type: establishing
    prompt: "Wide shot of abandoned research facility at dusk"
    duration: "5s"
    model: "veo-3-1"  # or "seedance-2" if using Studio
```

## Phase 2: Character Reference Generation (Venice API)

### The 4-angle system

For character consistency across scenes, generate **4 angles per character**:

1. **Front** — Primary reference face
2. **Three-quarter** — Slight turn, shows depth
3. **Profile** — 90-degree side view
4. **Full body** — Shows costume, proportions

```python
# Using venice-image-generate API
angles = ["front", "three-quarter", "profile", "full-body"]
for angle in angles:
    generate_image(
        prompt=f"{character_desc}, {angle} view, studio lighting, clean background",
        aspect_ratio="1:1",
        resolution="1024x1024"
    )
```

### Why 4 angles matter

- **R2V models** (Kling O3, Seedance 2.0) use these to maintain facial structure across motion
- **Different angles** prevent the model from "forgetting" what the character looks like when they turn
- **Profile shots** are critical for walking/running scenes

## Phase 3: Video Shot Generation (Venice API)

### Multi-shot per clip

Modern models (Kling O3, Seedance 2.0) can pack **multiple camera angles into a single generation**:

```json
{
  "model": "kling-o3-pro-image-to-video",
  "prompt": "Wide establishing shot of a cyberpunk alley. Cut to medium shot of @Element1 walking forward, determined expression. Cut to close-up of @Element1's eyes glowing with data.",
  "elements": [
    {
      "frontal_image_url": "https://.../srikhav-front.png",
      "reference_image_urls": [
        "https://.../srikhav-3q.png",
        "https://.../srikhav-profile.png"
      ]
    }
  ],
  "duration": "12s",
  "aspect_ratio": "16:9",
  "audio": true
}
```

### The "no music, only sound effects" rule

When generating clips via API:
- **Set `audio: true`** but prompt with `"no music, only ambient sound effects"`
- This gives you clean audio stems for later mixing
- Music added in Studio will be unified and smooth
- Music baked into each clip = choppy cuts and clashing tracks

**Exception:** Sora 2 Pro is **silent-only** — it accepts `audio: true` on `/video/quote` but rejects it on `/video/queue`. Generate Sora clips without the `audio` field, then layer music/SFX in post.

### Parallel queuing

The Venice API supports **simultaneous queue submissions**. Generate 5-10 shots at once while you work on the next scene's treatment.

```python
# Queue all shots for Scene 1 simultaneously
for shot in scene_1_shots:
    queue_video(shot)  # Each returns a queue_id immediately
```

## Phase 4: Audio Generation (Venice API)

### Music with timing cues

```json
{
  "model": "elevenlabs-music-v2",
  "prompt": "Cinematic orchestral score. First 10 seconds: soft and mysterious. Next 20 seconds: intense buildup. Following 10 seconds: calm tension. Final 15 seconds: climactic crescendo.",
  "duration": "65s",
  "instrumental": true
}
```

### Sound effects

Generate individual SFX clips for key moments:
- Explosions
- Footsteps
- Door creaks
- Ambient room tone

These are layered in Studio's Movie Editor or mixed locally with ffmpeg.

## Phase 5: Final Assembly (Venice Studio Web)

### Import assets

1. Go to `venice.ai/studio`
2. Open **Library** tab
3. Upload/download all generated clips, images, audio files

### Movie Editor workflow

1. **Create project** → Set target aspect ratio (9:16 for vertical, 16:9 for cinematic)
2. **Drag clips to timeline** → Arrange in shot order
3. **Trim** → Cut dead frames at start/end of each clip
4. **Add transitions** → Fade, blur, cross-dissolve between shots
5. **Sync audio** → Drag music track under video, align hits with action
6. **Add title cards** → Text overlay for title, credits
7. **Export** → MP4 download

### Seedance 2.0 for cinematic shots

If API models didn't nail a particular shot:
1. Go to **Video** tab in Studio
2. Select **Seedance 2.0 R2V**
3. Upload your 4-angle character refs
4. Craft the prompt manually (often better than API for fine-tuned control)
5. Generate and download
6. Swap into Movie Editor timeline

## Phase 6: Iterative Refinement (Hybrid Loop)

The power of this workflow is the **back-and-forth**:

| Issue | Fix Location | Action |
|---|---|---|
| "Character looks wrong in shot 7" | Telegram → Hermes | Regenerate with adjusted refs via API |
| "Transition between shots 3-4 is jarring" | Venice Studio | Adjust in Movie Editor |
| "Need a wider establishing shot" | Venice Studio | Seedance 2.0 R2V with scene refs |
| "Music doesn't hit at the right moment" | Venice Studio | Slide audio track, adjust fades |
| "Logo on character's forehead is wrong" | Telegram → Hermes | QA agent rejects, regenerates via API |

## Pitfalls

- **Trying to do everything in the API.** The Venice API cannot edit timelines. If a user asks for "add a transition" or "cut this clip," they must use Studio's Movie Editor.
- **Seedance 2.0 FOMO.** Seedance 2.0 produces the best cinematic results but is UI-only. Accept this limitation — use Kling O3 or Veo via API for bulk, then swap in Seedance shots for hero moments.
- **Missing the 24h download expiry.** API `download_url` expires in 24 hours. Download immediately or re-queue.
- **Inconsistent aspect ratios.** Mixing 9:16 and 16:9 clips in the same project requires Studio cropping or letterboxing. Generate everything in the target ratio from the start.
- **Forgetting the "no music" rule.** If clips have baked-in music, Studio assembly becomes a nightmare. Always generate with "sound effects only."
- **Not using reference images.** Blind text-to-video without character refs = inconsistent faces across shots. Always generate 4-angle refs first.
- **Overloading a single prompt.** While multi-shot prompts work, packing 5+ camera changes into 5 seconds confuses the model. Keep it to 2-3 cuts per 12-15 second clip.

## Reference Files

- `references/venice-studio-vs-api.md` — Full feature matrix (from `venice-video` skill)
- `references/micro-drama-notes.md` — Specific techniques from Venice tutorial videos

## Related Skills

- `venice-video` — Async video generation API (quote/queue/retrieve/complete)
- `venice-image-generate` — Character reference image generation
- `venice-image-edit` — Angle changes, face fixes, background removal
- `venice-audio-music` — Soundtrack generation with timing
- `venice-audio-speech` — Voiceover / TTS
- `venice-chat` — Treatment and script generation
- `venice-models` — Model selection and capability probing

## External Resources

- Venice Studio: `https://venice.ai/studio`
- Venice Video Harness (IDE workspace): `https://github.com/jordanUrbs/venice-video-harness`
- Venice Video Model Routing Skill: `https://github.com/jordanurbs/venice-video-model-routing`