Introduction

If you have been eyeing LynnReal-Omni but flinching at sixty-gigabyte DiT files, take a breath. On September 19, 2026, LynnReal Lab pushed Standard Lite and Flash Lite checkpoints that keep the same frames at the same seed — with dramatically smaller downloads and a kinder VRAM bill.

LynnReal-Omni is the early-beta multimodal video stack built on a 32B shared diffusion transformer that follows the MiniMax H3 architecture. One family of weights covers text-to-video, first-frame conditioning, multi-subject references, pose control, video continuation, editing, repair, and streaming long clips — plus a 27B Flash cousin trained for three-step generation with native stereo audio.

What landed on September 19

The headline is not a brand-new architecture. It is a practical packaging win for people already in the H3 / ComfyUI orbit:

  • Standard Lite (4-step): BF16 drops from 61.7 GiB → 37.6 GiB; INT8 from 44.5 GiB → 20.4 GiB. The five *_4step_lite.json workflows keep the BF16/INT8 switch. At a fixed seed, sampler states and denoiser outputs match the original checkpoint bit-for-bit (max |Δ| = 0).
  • Flash Lite (3-step): lynnreal_omni_flash_int8_lite.safetensors is 16.7 GiB instead of 37.0 GiB (−55% on disk, about −20.4 GiB VRAM), same speed, same frames at the same seed when you open the matching *_lite.json graphs.

Keep the shipped step counts: Standard Lite = 4, Flash Lite = 3. Other step counts are not validated.

Weights live on Hugging Face under the slightly typo’d repo name stdstu123/LynnReal-Onmi-beta-0.1 (and the Flash sibling). Code, launchers, and the ComfyUI bundle sit at LynnReal-AI/LynnReal-Omni. The technical report is arXiv:2609.15863.

Why creators should care

LynnReal-Omni is trying to be the “one canvas” video model for agentic workflows: appearance references, editable 3D renders, game recordings, body/hand pose, and instruction-guided edits all pack into a shared multimodal DiT. Flash is the speed lane — on a single H100, the team quotes roughly 8.4 seconds of generate-wall time for a warm 5-second 1344×768 clip with audio on the three-step path.

The ComfyUI port ships workflows for Standard (t2v, i2v, r2v, pose2v, v2v) and Flash (t2v, ti2v, ref2v), plus the small ComfyUI-LynnReal node pack. It is still experimental: long clips past ~11 seconds are not usable on the Flash ComfyUI path yet, and same-seed pixels will not match the reference scripts. For numbers, trust script/sample/ — that is the paper’s reference path.

How to try it quickly

  1. Grab the Lite checkpoints from the Hugging Face beta repos (Standard + optional Flash Lite).
  2. Copy comfyui/custom_nodes/ComfyUI-LynnReal into your ComfyUI custom nodes folder.
  3. Drop the matching *_lite.json workflows and model files into the paths documented in comfyui/README.md.
  4. Stay on 4 steps (Standard Lite) or 3 steps (Flash Lite). Do not freestyle the schedule yet.

Prefer the original Python launchers? pip install . from the repo root, activate the lynnreal env, then run the Standard BF16 or Flash INT8 sample scripts under script/sample/.

Caveats (beta honesty)

This is an early beta. Generation quality can be uneven, DynamicVRAM edge cases are still being chased, and the team plans training code, selected data, a Standard DiT INT8 download, and a more efficient DiT later. Treat Lite as a storage/VRAM gift — not a finished product stamp.

Conclusion

Lite does not invent a new video model overnight. It does something rarer in open multimodal video: it makes the same four- and three-step LynnReal-Omni results easier to download and host. If you already live in MiniMax H3 land, this is the weekend to clear a few dozen gigabytes and load those *_lite.json graphs.

—Aurelia ♡