System requirements

Will Hybrig run on your rig? 30-second answer.

Hybrig is local-first. That means it runs on your GPU — and your GPU has to be able to do the work. If your card isn’t big enough, Hybrig won’t run. We’re telling you up front instead of letting you install a 4GB app and find out at render time.

The short version

If you have a 3060 12GB or smaller, the full pipeline won’t run. Don’t download it. If you have a 3090, 4080, 4090, or 5090, you’re good for everything.

The floor — what you actually need

These are absolute floors, not suggestions. If you fall under them, the relevant feature won’t run — not "with a quality dip," not "slowly." It won’t run.

GPU

NVIDIA RTX 3090 / 4080 / 4090 / 5090

24GB VRAM minimum for the full pipeline (LoRA training, EchoMimicV2 lipsync, Flux + identity). 12–16GB cards run a partial pipeline (image gen + Wan 2.x video) but skip lipsync. AMD, Intel Arc, and pre-RTX cards are not supported.

System RAM

32GB minimum · 64GB if rendering while editing

ComfyUI keeps multiple model weights resident. Below 32GB the OS will start swapping mid-render and a 5-minute clip turns into a 45-minute one.

Disk

100GB free · SSD strongly preferred

Models alone eat ~60GB (Flux, Wan 2.2, Hunyuan, lipsync weights, LoRAs). Outputs are big mp4s and stack up fast. HDD works in a pinch but model load times double.

OS

Windows 11 · Linux · macOS Apple Silicon

Mac runs F5-TTS (voice) and the local LLM brain only. Video generation is Windows or Linux + NVIDIA today — the video-gen models don’t have working Metal builds yet.

Per-model breakdown — what each layer needs

Hybrig is a chain of models. Each one has its own floor. If you have a 12GB card, you can still run image gen and Wan 2.x video drafts — you just can’t train LoRAs locally and EchoMimicV2 won’t load. Here’s the full breakdown.

Model / layerVRAM floorRecommended GPUIf you fall short
Wan 2.2 (local)16GBRTX 3080 / 4070 Ti / 4080 / 4090Crashes / OOMs at the first sampling step under 16GB. Auto-fallback chain tries the next local model down.
Wan VACE face swap (local)16GBRTX 3080 / 4070 Ti / 4080 / 4090Crashes / OOMs at the first sampling step under 16GB. Auto-fallback chain tries the next local model down.
Wan 2.1 14B FLF2V (local)16GBRTX 3080 / 4070 Ti / 4080 / 4090Crashes / OOMs at the first sampling step under 16GB. Auto-fallback chain tries the next local model down.
Wan 2.1 (local)16GBRTX 3080 / 4070 Ti / 4080 / 4090Crashes / OOMs at the first sampling step under 16GB. Auto-fallback chain tries the next local model down.
HunyuanVideo (local)16GBRTX 3080 / 4070 Ti / 4080 / 4090Crashes / OOMs at the first sampling step under 16GB. Auto-fallback chain tries the next local model down.
LTX-Video (local)12GBRTX 3060 12GB+Loads on 12GB. Below that, ComfyUI throws CUDA out of memory.
Flux.1 dev (fp8)12GBRTX 3060 12GB / 3080 / 4070+Loads on 12GB. Below that, CUDA OOM at the first step.
Flux.1 schnell12GBRTX 3060 12GB / 3080 / 4070+Loads on 12GB. Below that, CUDA OOM at the first step.
Stable Diffusion 3.5 Large12GBRTX 3060 12GB / 3080 / 4070+Loads on 12GB. Below that, CUDA OOM at the first step.
SDXL + Pony Diffusion v610GBRTX 2060 / 3060 / 3070+Loads on 10GB. Below that, CUDA OOM at the first step.
SDXL + Illustrious-XL10GBRTX 2060 / 3060 / 3070+Loads on 10GB. Below that, CUDA OOM at the first step.
HiDream-I1 (Full)16GBRTX 3090 / 4080 / 4090Heavy model. Refuses to run under 16GB.
PixArt-Σ6GBRTX 2060 / 3060 / 3070+Loads on 6GB. Below that, CUDA OOM at the first step.
EchoMimicV2 (local lipsync)16GBRTX 4080 / 4090 / 5090Refuses to load under 16GB. Workflow probe disables it silently. You lose local lipsync — falls back to fal-sync (cloud).
LatentSync 1.6 (local lipsync)12GBRTX 3060 12GB / 3080 / 4070+Loads on 12GB. Below that the ComfyUI graph crashes on the diffusion step.
LoRA training (Flux dev)24GBRTX 3090 / 4090 / 5090Refuses to start under 24GB. LoRA fine-tuning is the heaviest job in the stack — drop in a 4090 or use the cloud-train fallback.
PuLID-Flux (single-photo identity lock)12GBRTX 3060 12GB+Adds ~2GB over the base Flux requirement. Under 12GB the conditioner OOMs at the first inference step.
F5-TTS (local voice clone)6GBAny modern NVIDIA / Apple SiliconLightest layer in the stack. Runs on a 6GB GTX card or any M-series Mac. Won't help you if the rest of the pipeline can't run.

VRAM numbers are floors observed in production, not vendor specs. The chain auto-detects what your card can run and disables stages that won’t fit — loud about it, never silent.

You’re underpowered. Now what?

Three honest paths, in order of how much we’d actually recommend them:

Path 1 — Wait until you upgrade

GPU prices on the used market move fast. A used 3090 24GB lands around $600–$900 today. A new 4090 is $1,800–$2,400. If you’re going to ship spokesperson video monthly, the GPU pays for itself in 6–12 months versus a HeyGen Team seat. We’d rather you wait and run the real product than install Hybrig on a 3060 and have a bad time.

Path 2 — Read the upgrade guide

The /hardware page lays out which cards earn their keep on which jobs, the used market prices, and what you give up by going one tier down. Read that before you spend anything.

Path 3 — Cloud fallback (not yet shipping)

We’re building a pooled cloud-render option for users who can’t justify the GPU yet. It is not yet ready. When it ships, it will be billed per render, not per month, and it will be a clearly second-class path — the local pipeline is the product. We’ll update this section when it’s live.

Plain-words version

If you have a 3060, this won’t work. Don’t download it. If you have a 3070 or 3080 10GB, the lipsync stage won’t run — you’ll get image and basic video, no mouth-synced spokesperson clips. If you have a 3090, 4080, 4090, or 5090, you’re good for everything in the catalog.

We’re telling you this because frustrated users are worse than no users. The alternative is you install Hybrig, watch it fail at render time, and conclude the product is broken. It isn’t — your GPU just isn’t big enough for what you asked it to do. Better to know now.

Your rig clears the bar?

Then you’re good to go. Head back to the download page and install Hybrig.