MiniMax H3 Ref2VA + LightX2V 8-step 768p (merged transformer)
This repository contains the premerged Ref2VA transformer only for native MiniMax H3 reference-to-audio-video generation. It combines the official MiniMax H3 Ref2VA BF16 transformer with LightX2V's minimax_h3_ref2v_turbo_8step_v1.0_768p_bf16.safetensors LoRA.
It must be assembled with the unchanged processor, tokenizer, text encoder, visual VAE, and audio VAE from the pinned MiniMaxAI/MiniMax-H3 Ref2VA partition. It is intended for 8-denoiser-evaluation, 768p ref2va inference. This is not a standalone LoRA and is not the Diffusers-format checkpoint.
Provenance
| Item | Value |
|---|---|
| Base model | MiniMaxAI/MiniMax-H3, native Ref2VA checkpoint |
| Base revision | 42ed227ee7df40d41602854ae760620d6eb651fe |
| LoRA source | lightx2v/Minimax-h3-Turbo |
| LoRA file | minimax_h3_ref2v_turbo_8step_v1.0_768p_bf16.safetensors |
| LoRA SHA-256 | 9bac880b1a5d7ac052171cf6cce769f0cceaaa42ffa51de4b8e41143a2bdd2d2 |
| LoRA rank / alpha | 128 / 8 |
| Effective merge scale | 8 / 128 = 0.0625 |
| Merged native tensors | 208 targets from 624 LoRA tensors |
| Weight precision | BF16 |
The adapter was converted from the Diffusers key layout to the native MiniMax H3 layout during the merge. Split Q/K/V deltas were interleaved into native per-head fused QKV tensors, and the Diffusers SwiGLU value/gate order was converted back to the native gate/value order. The machine-readable merge manifest is included as nauknauk_ref2va_merge.json.
The LoRA release and recommended 8-step 768p settings are documented in LightX2V discussion #51.
Assemble and deploy with SGLang
Download the pinned base partition, then overlay this repository's transformer:
hf download MiniMaxAI/MiniMax-H3 \
--revision 42ed227ee7df40d41602854ae760620d6eb651fe \
--include "model_index.json" "Ref2VA/*" \
--local-dir MiniMax-H3
hf download NaukNauk/minimax-h3-turbo-ref2va-merged \
--local-dir MiniMax-H3/Ref2VA
The checkpoint was validated with SGLang 0.5.17 on 30 reference-video cases at native 768p using four H100 80GB GPUs, tensor parallelism 4, Ulysses degree 1, and no CPU offload. All 30 cases completed, including inputs with six reference images.
CUDA_VISIBLE_DEVICES=0,1,2,3 sglang serve \
--model-type diffusion \
--backend sglang \
--model-path MiniMax-H3 \
--pipeline-class-name MiniMaxH3Pipeline \
--model-variant ref2va \
--num-gpus 4 \
--tp-size 4 \
--ulysses-degree 1 \
--performance-mode speed \
--use-fsdp-inference false \
--text-encoder-cpu-offload false \
--vae-cpu-offload false \
--dit-cpu-offload false \
--encoder-parallel auto \
--enable-torch-compile false \
--host 0.0.0.0 \
--port 30010
Recommended generation settings:
num_inference_steps: 9 in native SGLang, which includes the terminal zero and therefore executes 8 denoiser evaluations- video
flow_shift: 12 - audio
flow_shift: 3 - output short edge: 768 pixels
- frame rate: 24 FPS
The exact request schema depends on the SGLang release; consult the current SGLang MiniMax H3 API documentation.
License and use
The merged checkpoint remains subject to the MiniMax H3 Community License Agreement. The LightX2V repository is published under Apache-2.0. Users are responsible for complying with both upstream terms and applicable laws.
This repository is a convenience artifact prepared by NaukNauk. It is not an official MiniMax or LightX2V release. Repository ID: NaukNauk/minimax-h3-turbo-ref2va-merged.