character sheet AI
character consistency
n8n AI video
image-to-video
UGC Shorts automation
Higgsfield templates
Most “AI video workflow” GitHub repos stop at script → MP4. Creators who ship trending Shorts, UGC ads, and micro-dramas do something else: they lock a character sheet, write a scenario beat-by-beat, force image-to-video with the same face reference for character consistency, then automate the boring parts with an n8n AI video pipeline or a template suite like Higgsfield.
This guide lists AI film production playbooks and automations published or actively updated in the last six months (April–October 2026). Every entry explains the film type, tools, how character consistency is handled, how scenarios are built, and what sample or template you get — so you can copy a real production stack, not just a one-click generator.
What this list is (and is not)
This article is for creators who want the full creative operating system: trend format → scenario → character bible → stills → motion → captions → publish — with enough detail to copy the process.
Included
- Playbooks with character sheets / visual bibles
- Consistency methods (LoRA, PuLID, Soul ID, first/last frame)
- Scenario / shot-list templates
- n8n JSON workflows and Seedance / Veo / Kling nodes
- Template products (Higgsfield Viral Presets, Supercomputer)
- Repos created or materially updated Apr–Oct 2026
Excluded on purpose
- Black-box “topic → final MP4” agents with no character bible
- Older generic Remotion/script factories with no consistency docs
- Repos last touched before April 2026 (unless still maintained — noted)
- Cracked / license-bypass tooling
Freshness rule: Primary entries were created or meaningfully updated between 22 April 2026 and 22 October 2026. Dates are from public GitHub metadata and platform changelogs. Always re-check a repo’s updated_at before you invest GPU time.
How people actually shoot trending AI films

Across Shorts, Reels, TikTok, UGC ads, and vertical micro-drama, the winning AI film production pattern is the same:
- Pick a format first (viral preset, UGC selfie ad, lifestyle loop, or narrative short) — not a random prompt.
- Lock identity assets before motion: face plate, turnaround sheet, wardrobe, location plates.
- Write a scenario as beats (2–3 seconds each), not as a novel.
- Generate still keyframes (image models) that match the sheet.
- Animate with image-to-video (Veo, Kling, Seedance, Wan…) using the same reference.
- Chain frames: last frame of clip A becomes the first frame of clip B.
- Automate the queue with n8n / Sheets / Telegram review — humans still approve faces.
Template platforms like Higgsfield Trending Templates collapse steps 1–5 into a preset. Serious narrative work still uses the full bible (see playbooks below).
How creators build character sheets

A character sheet AI asset is the non-negotiable source of truth for face, silhouette, hair, and costume. Typical 2026 recipe:
1) Pose grid → multi-angle sheet
Feed an OpenPose / skeleton grid into ComfyUI (FLUX or SDXL). Generate front, ¾, profile, and back views in one sheet. Upscale, then crop faces and full poses.
2) Expression pack
Generate neutral, smile, serious, surprise — same lighting. Keep 10–15 clean crops.
3) Identity lock
- FLUX LoRA (FluxGym / ai-toolkit) with a unique trigger word (e.g.
maya_ar) — best for long projects. - PuLID / PuLID-Flux — face lock per generation without full LoRA training.
- Platform Soul ID / Elements (Higgsfield, Kling Element Library) — train once, reuse across models.
4) Canonical reference rule
Pick one hero face plate for the whole project. Do not rotate between different “good” faces mid-shoot — that is the #1 cause of identity drift.
Minimum character bible checklist
- Neutral front face (even light)
- Full-body silhouette
- Profile or ¾ view
- Wardrobe / costume plate
- Signature prop (if any)
- Voice note / TTS profile (for dialogue shorts)
- Forbidden variants (no smile, no glasses, no costume change…)
How consistency is kept across shots

| Method | When to use | Tools (2026) |
|---|---|---|
| Character LoRA | Multi-episode series, complex wardrobe | FLUX LoRA, FluxGym, CharForge-style pipelines |
| PuLID / Face ID | Photosets + I2V without training a LoRA | ComfyUI PuLID-Flux, SciensOne generator |
| First + last frame | UGC ads, lifestyle Shorts | Veo 3.1, NanoBanana edit → video; Kling start/end |
| Frame chaining | Narrative shorts | Export last frame → next I2V start |
| Platform identity | Fast trending / short films in one suite | Higgsfield Soul ID, Kling Elements |
| QA gates | Production systems | Showrunner identity/prop/scale checks; DeepFace distance |
Community consensus in 2026: IP-Adapter alone is weak for identity. Treat it as style transfer. Prefer LoRA, PuLID, or platform Soul ID for faces.
How scenarios and shot lists are written

Creators who look “directed” do not paste a paragraph into a text-to-video box. They:
- Write a short treatment (logline, genre, tone).
- Break scenes into 2–3 second beats.
- For each beat, separate immutable elements (face, costume, location palette) from mutable ones (action, camera move, expression).
- Prompt physical blocking (“right corner of mouth lifts, eyes unchanged”) instead of vague emotion words.
- Run a 5–10 shot proof sequence before spending on the full film.
Playbooks like billpar/ai-cinematic-pipeline and FuFicFac/fiction-to-ai-video-production document this end-to-end.
Lighting, camera & effects in AI film prompts
Identity lock alone still looks “AI-generic” if every clip uses flat daylight and a static mid-shot. Production playbooks treat lighting, lens language, and effects as locked fields — the same way they lock the face plate.

1) Lighting bible (copy into every still prompt)
- Key light — direction (camera-left / right), hardness (softbox vs hard), intensity relative to fill (e.g. 2:1).
- Fill — how much shadow detail you keep; “no fill” = noir; “strong fill” = UGC beauty.
- Rim / hair light — separates subject from background; critical for dark wardrobes on dark sets.
- Color temperature — one Kelvin band per scene (e.g. 3200K tungsten interior, 5600K daylight exterior). Do not mix mid-project.
- Practicals — neon signs, window spill, product LED — name them once in the location plate.
Write lighting as immutable prompt tokens: soft key camera-left 2:1, cool rim, 5600K, no mixed neon. When the face drifts, creators often “fixed” identity but changed light — models treat that as a new person.

2) Camera language (per beat, not per paragraph)
| Field | What to specify | Why it matters for I2V |
|---|---|---|
| Shot size | ECU / CU / MCU / MS / WS | Stops the model from zooming randomly between clips |
| Focal length feel | 24mm wide, 35mm documentary, 85mm portrait | Keeps face proportions stable across the cut |
| Camera move | static / slow push-in / orbit / handheld / crane | One move per 2–3s beat; stack moves in edit, not in one prompt |
| Depth of field | shallow bokeh vs deep focus | Product UGC usually shallow; narrative establishing = deep |
| Angle | eye-level / low / high / Dutch | Dutch and low angles amplify “cinematic” without new wardrobe |
3) Effects & finishing (one look per episode)
- In-camera / prompt effects: anamorphic streak, volumetric haze, lens flare (sparingly), motion blur on pans only.
- Post LUT / grade: lock one grade (teal-orange short, clean UGC, Ghibli soft) in FFmpeg or the platform’s color node — not a new look every ASIN.
- Grain & halation: add in finishing so image models stay sharp for identity lock.
- Speed ramps / whip pans: edit-side transitions between chained clips; do not ask the video model for a whip + dialogue + product insert in one generation.
Prompt template (drop into each beat)
[immutable: character trigger + wardrobe + location + lighting bible] · [shot: MCU 85mm eye-level] · [move: slow push-in] · [action: lifts product to chest height] · [effects: soft haze, no flare] · [grade: clean UGC]
Playbooks such as billpar’s cinematic pipeline and FuFicFac’s fiction guide treat lighting and camera as part of the visual bible — the same discipline applies whether you use ComfyUI stills, Veo/Kling I2V, or Higgsfield Cinema Studio.
Methodology playbooks (last 6 months)
These repos teach how to produce — not only how to call an API.
billpar/ai-cinematic-pipeline Jun 2026 · updated Oct 2026
Film types: narrative shorts, episodic series.
Character sheets: locked 7-panel turnaround references; setting plates; prop plates; one voice profile per character.
Consistency: immutable vs mutable prompt fields; frame chaining (last frame → next start); never regenerate the bible mid-project.
Scenario: script breakdown → 2–3s beats → one prompt per beat → audio post as separate tracks.
Tools named: tool-agnostic methodology with platform notes in docs; Python helper to turn a breakdown into a shot list.
Samples / templates: docs/ methodology + templates/ for sheets, shot lists, prompts.
FuFicFac/fiction-to-ai-video-production Active 2026 playbook
Film types: fiction / narrative shorts built from a visual bible first.
Character sheets: detailed character + location + prop bible before any stills or motion.
Consistency: still references first; fix identity before video; continuity rules + negative prompts.
Scenario: treatment → bible → reference stills → 5–10 shot proof sequence → full cut.
Tools: tool-agnostic methodology (works with Kling / Veo / Runway-class I2V).
Samples: written production guide with copy-paste bible examples.
ndpvt-web/capy-video-gen-skill Updated Sep 2026
Film types: multi-shot narrative with measured face consistency.
Character sheets: front / side / back portraits extracted from the script.
Consistency: keyframe face-lock then Veo/Sora I2V; DeepFace distance metrics (~70% improvement in their tests).
Scenario: idea-to-script agent or screenplay → storyboard → clips → FFmpeg assembly.
Tools: Gemini image, Veo 3.1 / Sora 2, FFmpeg.
Samples: runnable skill pipeline + experiment tables in the README.
emezac/showrunner Jul 2026 · pushed Aug 2026
Film types: consistency-first short films with QA gates.
Character sheets: canonical visual bible for characters, props, locations, scale, physics.
Consistency: storyboard visual QA (identity / prop / scale / physics); bounded auto-repair; operator overrides.
Scenario: prompt → validated screenplay → storyboard contract → video → assemble MP4.
Tools: Ruby app orchestrating generative providers; experimental but extremely explicit contracts.
Samples: full production pipeline with QA ledger.
sinmb79/22b-studio Apr 2026 · updated Aug 2026
Film types: Ghibli-style animated shorts (scenario → finished clip).
Character sheets: stage 2 of a 7-stage pipeline (front / side / full-body sheet).
Consistency: IP-Adapter optional; human review after cheap sketch stage before full render.
Scenario: single scenario input → mood board → sheet → background → sketch → full render → AnimateDiff → upscale.
Tools: n8n-first + ComfyUI SDXL + Ghibli LoRA + FFmpeg + Real-ESRGAN.
Samples: importable 22b-studio-ghibli.json plus per-stage ComfyUI graphs.
ComfyUI character-sheet & identity lock (last 6 months)

SciensOne/comfyui-workflow-generator Jun 2026
What it is: parametric batch generator for consistent-character photosets + matching image-to-video.
Character sheets: describe characters / poses / locations in data files → emit ComfyUI JSON batches (including 10×10 pose grids).
Consistency: PuLID identity lock on SDXL and Flux/GGUF routes; dual-face masked pairs; I2V from locked first frames; optional MMAudio foley.
Scenario: config-driven scene lists rather than one-off prompts.
Samples: example workflows + docs in-repo.
Mickmumpitz Flux Consistent Characters (community mirrors)
Still the most-cited free pattern for sheet → crops → FLUX LoRA. Mirrors such as eusouia/Free-Streaming-Consistent-Characters package the ComfyUI graphs; always credit Mickmumpitz’s tutorial for the method. Use it when you need a durable LoRA for long affiliate series.
n8n automation pipelines (last 6 months)

These are the n8n AI video workflows creators actually import — Sheets as a queue, Telegram for review, Veo / Kling / Seedance for motion. Pair this with the storytelling habits in our 7-step creator storytelling guide so automation does not flatten your narrative.
madebysaira/n8n-ai-creator-pipelines Jul 2026
Film types: commercial / client campaigns, multi-model A/B.
Flow: brief intake → prompt variations → multi-model batch (Kling / Veo / Runway / Luma / Higgsfield) → client review → final render → delivery → Shorts repurpose.
Consistency: human art-direction gate before final render; parallel models compared on the same prompt.
Samples: importable JSON for each stage; real jewellery campaign walkthrough in the README.
kkuxb/n8n-nodes-seedance2 Apr 2026
Film types: Seedance 2.0 text-to-video, first-frame I2V, first+last frame, multimodal reference video.
Consistency: first/last frame and multi-reference modes keep identity + product locked.
Scenario: node operations for create / poll / download; pair with Sheets for shot queues.
Samples: documented create-and-wait and Wait-loop poll patterns.
AtlasCloudAI/n8n-nodes-atlascloud maintained 2026 · Seedance 2 / Veo 3.1 / Kling 3
Community n8n nodes for 300+ models including Veo 3.1, Kling 3, Seedance 2, Nano Banana Pro, Sora 2, Wan. Importable async video workflows live under workflows/. Use when you want one credential and many model backends.
Official n8n.com templates (Veo 3.1 + NanoBanana)
Still useful blueprints even when the gallery “last update” label lags:
- #11594 Consistent character videos — location/pose roulette, start+end frames, Veo 3.1, auto TikTok/IG captions.
- #12752 UGC via Sheets — product + character + background → NanoBanana Pro → Veo 3.1.
- #13015 Ad clones — analyze a winning ad → rebuild scenes with Kling → publish.
Higgsfield-style template platforms

If GitHub playbooks are the “film school,” Higgsfield is the “preset pack + studio.” Useful when you need volume and trend formats fast.
| Piece | Role in the pipeline | Consistency |
|---|---|---|
| Viral Presets / Trending Templates | One photo → trend motion/style (mukbang, comic, roller coaster, magazine…) | Source face photo |
| Short film pipeline | Popcorn storyboard → Cinema Studio → Seed Audio → Supercomputer | Soul ID trained once, reused across Kling / Veo / Seedance |
| Viral short-film case | Character/location sheets first → Claude skill scenes → Seedance 2.0 | Fixed asset sheets before any motion |
| Supercomputer automation | Scheduled batches + ChatGPT/Claude MCP; API for custom products | Same Soul ID / brand brief in memory |
Workflows by trending film format
| Format | Best starting stack | Consistency method | Automation |
|---|---|---|---|
| Viral photo→effect Shorts | Higgsfield Viral Preset | 1 clear face photo | Supercomputer schedule |
| UGC product selfie ad | n8n #12752 / UGC agent + Veo first/last | Person + product refs | Google Sheets queue |
| Lifestyle influencer loop | n8n #11594 (poses × locations) | Start/end frame identity | Auto captions + publish |
| Vertical micro-drama | FuFicFac bible + Capy / Kling/Veo I2V | Bible + turnaround + chain | Optional n8n shot queue |
| Cinematic mini short | billpar pipeline / Showrunner / Higgsfield Soul ID | Sheet → LoRA or Soul ID | Supercomputer / showrunner QA |
| Animated stylized short | 22B Studio (n8n + ComfyUI) | Character sheet stage + review gates | Full 7-stage n8n graph |
| Client commercial A/B | madebysaira pipelines | Same prompt across models | Telegram review + delivery |
| Ad creative clones | n8n #13015 (NanoBanana + Kling) | Structure from winning ad | Airtable + Blotato |
Recommended tool stack (October 2026)

Identity & stills
- ComfyUI + FLUX / SDXL
- PuLID-Flux or character LoRA
- NanoBanana / GPT Image for clean edits and contact sheets
Motion
- Google Veo 3.1 (strong first/last frame)
- Kling 3 / Element Library
- ByteDance Seedance 2.0 (native audio options)
- Wan / LTX for self-hosted experiments
Orchestration
- n8n (self-hosted preferred for media binaries)
- Google Sheets / Airtable as the shot queue
- Telegram or Slack for human face QA
- FFmpeg for stitch, loudness, captions burn-in
Template acceleration
- Higgsfield Viral Presets + Soul ID + Supercomputer
Why this matters for Amazon affiliates

Affiliate growth in 2026 is increasingly video-first: Shorts reviews, UGC-style product demos, and lifestyle loops that stay on-brand — the same shift covered in our social commerce playbook for Amazon brands. The same character consistency systems above let you:
- Keep one presenter face across dozens of ASIN videos (trust + recognition) — close to the “trust triangle” dynamics in Trust Triangle Affiliate Sales.
- Batch UGC creatives for brand partners without reshooting, then distribute via Instagram and Shorts using patterns from the Instagram affiliate marketing blueprint.
- Repurpose one approved short into 9:16 / 1:1 / 16:9 for every channel while your commission stack stays organized (see repeatable product stack).
If you promote Amazon products through partnerships (not only classic Associates links), compare options in best Amazon affiliate marketing platforms, then use AlgoRift to discover brands, manage commissions, and track earnings while your video factory runs in n8n. Creators who already publish can layer AlgoRift Affiliates on top of the same character-locked content system.
FAQ
What is the single most important step for character consistency?
Lock a visual bible (face plate + turnaround) before any video generation, then use image-to-video with the same canonical reference every shot. Text-to-video invents a new face each time.
Should I train a LoRA or use PuLID / Soul ID?
Train a LoRA for long series and complex costumes. Use PuLID or platform Soul ID for faster photoset + I2V work. Many teams use Soul ID for volume Shorts and a LoRA for flagship narrative episodes.
Is n8n required?
No. You can run ComfyUI and Higgsfield manually. n8n becomes valuable when you have a queue (Sheets rows), need multi-model batches, client review, or daily publishing.
Why exclude older script→MP4 repos?
They optimize for autonomy, not for trending formats, character sheets, or continuity. This guide targets the workflows creators actually document for consistent faces and scenarios.
How often should this list be refreshed?
Every 60–90 days. Video models and n8n nodes move quickly; always verify GitHub pushed_at and changelog notes before adopting a stack.
Do lighting and camera need to be in the character bible?
Yes. Lock key/fill/rim, color temperature, and a default lens language (shot size + focal length feel) in the same bible as the face plate. Changing light mid-project is a common cause of “identity drift” even when the face reference is correct.
Build the video. Monetize the audience.
Use the AI video workflows above to ship consistent Shorts and UGC — then connect with Amazon brands that pay for real performance on AlgoRift.
Editorial note: tool names and repo dates were verified against public GitHub and vendor pages as of 22 October 2026. Third-party projects change; AlgoRift is not affiliated with the listed GitHub authors or model vendors.