ComfyDirector is an open-source playground on top of ComfyUI. Give it a prompt, pick a workflow, and try ideas without learning node graphs, wiring models by hand, or living inside a giant workflow mess.
Switch between cinematic, widescreen, portrait, and social crops to see how framing changes the result. It is a quick way to test composition before you commit to a workflow.
ComfyDirector started as a way to mess with AI video without constantly wrestling with raw node graphs.
It wraps ComfyUI in a simpler layer so you can throw in a prompt, try different templates, and compare results without learning every workflow detail first. The point is not to hide the mess completely. The point is to make it easier to play with.
A lighter way to mess around with prompts, templates, framing, and workflows without learning raw ComfyUI first.
The same prompt can look completely different at 2.35:1, 16:9, 9:16, or 4:5. ComfyDirector lets you compare those shapes early instead of treating aspect ratio like an afterthought.
Start with a rough idea and let the built-in LLM turn it into scenes, shot directions, and prompts you can actually test.
See key images and scene beats before you commit to a heavier video run.
Try camera presets and shot variations without touching a raw ComfyUI graph.
Run it on your own GPU or move the stack to cloud hardware when local VRAM is not enough. RunComfy support is there too, but still experimental.
Swap between cinematic, widescreen, portrait, and social framing to see how the same idea changes.
Your tests do not disappear into random output folders. Prompts, images, videos, and variants stay together.
Make a rough cut inside the app. Reorder clips, trim them, and stitch together something watchable without leaving the experiment.
Write the idea in plain English. No node graph needed.
AI turns it into scenes you can review, tweak, and approve before the heavy stuff runs.
Pick a template and let ComfyDirector run the matching ComfyUI workflow behind the scenes.
Clips and audio can be stitched into a simple final video.
Built for messing around: rerun stages, swap templates, tweak prompts, change seeds, and keep going until something interesting shows up.
8 Gen 2 templates across 5 GPU tiers — from lighter 8GB experiments to 32GB+ LTX 2.3 cinematic and lip-sync workflows.
Fast text-to-video with Wan 2.2 5B. Sound design via MMAudio. Perfect for quick experiments on modest GPUs.
Synchronized audio and video in a single pass with LTX-2. Camera control, face enhancement, and LoRA support built in.
Dual-expert MoE architecture delivers the highest quality text-to-video. LoRA support for style customization.
Specialized anime pipeline with curated LoRAs: retro 90s, modern HD, and Kawajiri styles. Built on the 14B MoE backbone.
The complete cinematic pipeline. Hybrid video (S2V lip-sync + LTX-2 motion), TTS narration, multi-character support, face enhancement, and opt-in image refinement before video.
Experimental LTX 2.3 native audio+video. Keeps the Qwen image and camera stages, then upgrades the final pass to newer synchronized dialogue and ambient audio generation.
Direct LTX 2.3 text-to-video with synchronized audio. Best when you want the newest cinematic backend without the full multi-stage image and camera pipeline.
Audio-driven LTX 2.3 talking shots from one scene image plus uploaded or generated dialogue. Built for cinematic talking-head clips with storyboard image fallback.
Plus 6 legacy templates (Gen 1) still available for backward compatibility.
ComfyDirector runs locally if you have enough GPU headroom. The lighter templates start
around 8GB VRAM, while the newest LTX 2.3 routes start at 32GB and the direct
LTX Cinematic workflow is aimed at 48GB-class cards. Most of the
hands-on testing so far has been on an RTX 5090.
If your machine is smaller, cloud GPUs are still a valid way to play with it. Some
templates also support the RunComfy API, though that path is still
experimental.
ComfyDirector is an open-source side project built in spare time. If it saves you from wrestling with
raw ComfyUI just to test a prompt, Buy Me a Coffee helps cover the time and experiments behind it.
If you would rather not donate, starring the repo, opening issues, and contributing workflows or
templates are still extremely useful signals.