Video Generation · Open-source models in production

Production video generation with open-source models.

LTX-Video, Wan 2.1, HunyuanVideo, Mochi-1, CogVideoX - deployed, brand-tuned, and wired into your content pipeline. Render times that match a campaign calendar, not a feature-film schedule.

~30s
render for a 5s LTX-Video clip on H100
5
OS video models in our production rotation
1080p
default output, 4K available
18×
speedup vs prior pipeline (case study)
Models we deploy

The current open-source video lineup

We benchmark these per project. There is no single winner - the right model depends on duration, motion type, render budget, and how much prompt-adherence you need.

LTX-Video 2.3

Lightricks

Fastest production-quality OS video model. Best motion fidelity at <6s clips. Our default for branded b-roll.

CinematicFastestT2V + I2V
Alibaba

Wan 2.1

Alibaba

Strong physical realism and camera control. Great when motion accuracy matters more than render time.

RealismCamera control

HunyuanVideo

Tencent

13B params, top-tier prompt adherence. Slower renders, best when prompt-to-shot accuracy is critical.

Prompt fidelity13B
HuggingFace

Mochi-1

Genmo

Open-weight 10B model with strong cinematic motion. Good middle-ground for narrative shots.

10BMotion
CogVideo

CogVideoX

Tsinghua

Long-form (up to 10s) generation with consistent characters. Useful for storyboards and previz.

Long-formCharacter consistency
Model comparison

Render time, duration, fidelity

Reference numbers on a single H100. Real-world throughput depends on prompt, batching, and quantization.

ModelMax durationResolutionRender speed (H100)Best for
LTX-Video 2.3~6s768×512 → 1216×704~30s for 5s clipFast cinematic b-roll
Wan 2.1~5sUp to 720p~2 min for 5s clipPhysical realism, camera moves
HunyuanVideo~5s720p+~4 min for 5s clipPrompt-accurate single shots
Mochi-1~5.4s480p → 720p~3 min for 5s clipNarrative cinematic motion
CogVideoX~10s720p~6 min for 6s clipLong shots, character consistency
Production pipeline

From brief to MP4

Generation is one box in a five-box system. We build the other four boxes too - that's the difference between a demo and a pipeline.

STEP 01

Script / brief → shotlist

An LLM step converts an editorial brief into a shotlist (subject, camera move, mood, duration). Each shot becomes a structured generation job.

STEP 02

Brand LoRA + style controls

We train a style LoRA on your existing footage (color, framing, depth of field). Applied at inference so every clip lands on-brand.

STEP 03

Generation + upscale

Base model generates at native resolution; second pass upscales to 1080p/4K (Real-ESRGAN, Topaz). Frame interpolation to 60fps where needed.

STEP 04

Audio + delivery

Optional MusicGen / Stable Audio scoring, branded TTS voiceover, then MP4 delivered straight into your NLE bin via API.

Use cases

Where this lands

Marketing b-roll

Brand-styled cinematic clips at the pace of a campaign calendar.

Product visualisation

Concept videos for unreleased products without a film crew.

Social content engine

Daily reels / shorts at platform-native aspect ratios.

Storyboard / previz

Long-form CogVideoX shots for narrative previsualisation.

Have a video pipeline to scope?

Bring a brief, some reference footage, a target turnaround. We'll come back with a model + LoRA + render plan.

Scope a video pipeline
See case studies