A Short-Drama (微短剧) Production Playbook

Short-drama — the fast-moving episodic format popularised on vertical video platforms — imposes constraints that single-shot AI generation struggles with. You need the same lead character to look identical in episode 1 and episode 12. You need talking-head dialogue that feels emotionally present. You need to produce 20–40 distinct scenes per episode quickly, and you need visual consistency across interior/exterior sets, lighting moods, and supporting characters.

Most creators try to solve this by re-running the same prompt repeatedly and hoping for the best. That approach yields inconsistent faces, drifting costume colours, and mismatched backgrounds. The solution is a disciplined reference-lock system combined with a multi-step workflow that propagates your locked assets into every scene automatically. Floniks is built around exactly this pattern.

Before generating a single scene, invest time in creating a canonical character sheet. Use /ai-image to produce a neutral front-facing portrait of your lead: soft even lighting, plain background, no extreme expression. Note every detail that must remain stable — hair colour, eye colour, facial structure, clothing palette.

Save this image as your reference seed. In the /editor workflow you will wire it into every image node via the "reference image" input so the model treats it as a style anchor. Run 10 portrait variants using the same reference and discard any that drift. The survivors form your character canon. Do the same for each recurring supporting character. Spending 30 minutes here saves hours of frame-by-frame consistency fixes later.

Tip: describe your character in a reusable prompt segment (e.g., "young woman, straight black hair to shoulders, wearing a white linen shirt, warm olive skin tone") and save it as a named template inside the editor so it auto-populates every scene node.

Open /editor and create a multi-step workflow with this spine:

1. Input node — accepts the episode scene list (a JSON array of scene descriptions).
2. Image generation node (per scene) — receives the character reference plus the per-scene description. Use a loop or fan-out branching node to process all scenes from the input array without rebuilding the workflow per episode.
3. Upscaling node — pipe every output through an upscaling step to normalise resolution before export.
4. Output collector node — bundles all scene images with filenames keyed to episode/scene numbers.

With this pattern you submit the episode 3 scene list, hit Execute, and receive 30+ consistent scene images in one run. The character reference flows into every node automatically — you never copy-paste a prompt manually per scene. This is the core efficiency gain of workflow-based short-drama production versus one-at-a-time generation.

Dialogue scenes require more than a static image. Navigate to /ai-avatar and upload the canonical portrait you locked in Step 1 as your avatar base. Record or synthesise the dialogue audio for the scene, then attach it as the audio input. Floniks drives the avatar's lip sync and subtle facial animation from the audio waveform.

Key practices for believable avatar dialogue:

• Keep individual takes under 45 seconds; string multiple takes in post rather than running one long generation.
• Use a slight head-angle variation (not full profile) — frontal faces animate most reliably.
• Supply an emotion cue in the prompt field ("anxious, brows slightly raised") to nudge the expression generator.
• If the character wears a specific costume, ensure the avatar base image already shows that costume — swapping clothing mid-scene breaks character consistency.

For crowd or reaction shots where dialogue is not needed, reuse the static scene images from Step 2 rather than burning avatar credits.

A character who looks identical but stands in a living room that changes colour temperature every episode will still feel discontinuous to viewers. Build a location library alongside your character library:

• Generate 3–5 canonical versions of each recurring set (the apartment, the office, the café) using /ai-image with a consistent lighting descriptor.
• Lock those set images as background reference inputs in your workflow nodes just like you do for characters.
• Use the same time-of-day and colour-temperature language across episodes ("warm golden-hour interior, practical lamp on left, deep shadow on right").

When you need a new location for a single episode, create it, note its full prompt descriptor, and archive it. If the location reappears two episodes later you can regenerate it accurately from the archived prompt rather than guessing.

Once all scene images and avatar videos for an episode are exported, assemble them in your video editor of choice. Floniks handles generation; the cut lives in your timeline tool. Standard short-drama pacing guidance:

• Hook in the first 3 seconds: lead with a high-tension or visually striking scene image before revealing dialogue.
• Scene length: 4–8 seconds per scene image with a subtle pan or zoom applied in the timeline keeps vertical video viewers engaged.
• Dialogue pacing: avatar clips should drive the emotional beats; cut away to a scene image for reaction or context, then return to avatar for the response.
• Cliffhanger tail: end each episode with a scene image that poses a visual question — a door about to open, a letter held up, a tense expression. This drives episode continuation.

Batch-produce multiple episodes in one Floniks session by queuing episode 1, 2, and 3 scene lists simultaneously in the workflow. The generation runs in parallel and all outputs land in the same session for download.

Pitfall 1 — Changing the character prompt mid-series. Even a small word change ("short black hair" vs "black hair, ear-length") can shift the model's interpretation. Lock your character prompt in a saved template and never edit it mid-production.

Pitfall 2 — Ignoring aspect ratio. Short-drama is 9:16. Generate all assets in 9:16 from the start. Cropping a 16:9 scene to vertical almost always cuts off the character's head or feet.

Pitfall 3 — Over-generating unique sets. Viewers don't notice if the same background appears in episodes 1, 4, and 7. Reusing locations saves generation time and reinforces the world's coherence.

Pitfall 4 — Skipping the upscaling step. Avatar videos and AI images may have slightly different native resolutions. Running everything through a unified upscaling node in the workflow normalises resolution and sharpness before export, producing a cleaner final cut.