LynnReal Flash

Code: LynnReal-AI/LynnReal-Omni

A 42-block joint video/audio DiT with three denoiser evaluations. Supports text-to-video and native first/last keyframes. Middle blocks use spatial stride-two token selection with full-resolution residual restoration.

The complete transformer uses trained W8A8 projections: INT8 weights, per-token INT8 activations, INT32 accumulation, and BF16 outputs. Sensitive projections remain BF16. Packed weights and per-channel scales are stored without requantization in indexed Hugging Face safetensors shards.

Load using the release model.flash_int8.load_transformer or the launchers in script/sample/flash/int8. This custom packed format requires that loader; a generic Diffusers from_pretrained does not install the INT8 operators. Shared conditioning and codec components are included in this model bundle. See LICENSE and NOTICE for upstream terms.

⚑πŸ”₯ ComfyUI β€” the fastest way to try it!

This checkpoint also runs in ComfyUI, at its trained three steps, from three workflows that ship with the release: t2v, ti2v (one first frame) and ref2v (reference pictures), plus the Light VAE and the node pack that implements the token selection! On a single H100 80 GB at 1344Γ—768, warm, three measured runs per cell:

Task 5 s Β· generate 5 s Β· click-to-video 10 s Β· generate 10 s Β· click-to-video
Text β†’ video 8.4 s 12.2 s 22.0 s 29.1 s
First frame β†’ video 8.9 s 13.0 s 23.1 s 30.1 s
References β†’ video 9.5 s 13.2 s 24.2 s 31.1 s

A five-second 1344Γ—768 clip with native stereo audio in about eight and a half seconds on one card: three denoiser steps and both decoders included. generate is the Generate-wall convention (first denoiser forward to decoded frames), click-to-video is what you actually wait for! Videos longer than 11 seconds are not usable in that path yet β€” the accelerated path for long clips is still being fixed.

Everything it needs β€” workflows, node pack, weights and demo assets β€” is in πŸ€— stdstu123/LynnReal-Onmi-beta-0.1 Β· comfyui. The weights below are the original bundle the script/sample/ launchers load.

Downloads last month
181
Safetensors
Model size
28B params
Tensor type
F32
Β·
BF16
Β·
I8
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support