Open-H3-Regenerate-2K-IC

Regenerates a low-resolution MiniMax-H3 draft (768P, 1344×768) at 1080P, 2K (2560×1440) or 4K (3840×2160) while keeping the draft's faces, lip sync, lettering and layout. The draft stays in the context of MiniMax-H3's Ref2VA transformer (transformer_ref) as a reference video; an in-context (IC) LoRA on that transformer generates the high-resolution version of the same clip.

Project page · Code

Files

Path	Content
lora/open_h3_regenerate_2k_ic_lora.safetensors	IC LoRA for transformer_ref (FP32, rank 384, alpha 384, 312 linears)
lora/minimax_h3_hyperflow_8step_v1.0.safetensors	HyperFlow 8-step LoRA, unchanged
lora/minimax_h3_lms_v1.0_r64.safetensors	LMS Sharp v1.0 LoRA, unchanged
lora/lora_config.json	IC LoRA key format and how it is applied
latent_upsampler/	Author 3D-conv latent upscaler (start point of the up3 / up4 samplers)
vae_fast/	Fast H3 VAE decoder v2 (non-causal, 24 layers)
assets/	Silent audio latent (required)
base_model/	MiniMax-H3 Diffusers components (transformer_ref, VAE, Qwen3-VL text encoder, tokenizer, processor, scheduler), unmodified copy of revision 42ed227, with its LICENSE
infer/	Inference code (infer.py), download.py, merge_lora.py; the same code as the GitHub folder
config.json	Components, output sizes and sampler settings read by infer.py

base_model/ mirrors MiniMaxAI/MiniMax-H3 @ 42ed227 without changes (only the components the IC pipeline uses; the T2V transformer/, the audio VAE and the native FL2VA/ and Ref2VA/ folders are left out). infer/download.py fetches everything here (--base-from-minimax takes base_model/ from the official repository instead) plus LMS Sharp's [1025, 8] adaln_t_table, read by byte range from Comfy-Org/MiniMax-H3.

Usage

pip install -r infer/requirements.txt
python infer/download.py --out weights
cd infer
torchrun --nproc_per_node 8 infer.py --weights ../weights --input draft.mp4 --output out.mp4 --caption "..."

Option	Values
--resolution	2k 2560×1440 (default) · 1080p 1920×1080 · 4k 3840×2160 · 768p from a 384P / 512P / 576P draft (--lq)
--sampler	up3 3 steps (default) · up4 4 steps · hf8 8 steps
--sharp	LMS Sharp strength (default 1.0, 0 = off)
--vae	fast fast H3 VAE decoder (default) · standard released H3 VAE
--jobs	a jsonl of {"input", "output", "caption"}: several clips, each model loaded once

up3 and up4 first enlarge the draft with the latent upscaler, noise it to σ 0.835 / 0.923 and run the last 3 / 4 points of the HyperFlow 8-step grid; hf8 starts from noise and runs all 8. All use CFG 1. The transformer is sharded over 1, 2, 4, 7 or 8 GPUs (FSDP + Ulysses context parallelism).

LoRAs

infer.py merges the HyperFlow 8-step LoRA and then LMS Sharp into transformer_ref at load time, and applies the IC LoRA on top. For runtimes that load one LoRA per model, python merge_lora.py --weights ../weights writes the IC LoRA and Sharp as one exact LoRA (factors concatenated along the rank axis) plus Sharp's curve branches; HyperFlow is loaded separately, before it.

Prompting

Pass the MiniMax-H3 prompt that generated the draft as --caption, unchanged. No extra prompt engineering is needed; the caption may also be left empty.

License

MiniMax H3 Community License (see LICENSE and NOTICE; the European Union, the United Kingdom, the Republic of Korea and the United States of America are excluded). base_model/ is an unmodified copy under that License (base_model/LICENSE). Third-party components keep their own licenses: HyperFlow (MiniMax H3 Community License), latent upscaler (Apache-2.0 weights, MIT code), fast VAE decoder (MiniMax H3 Community License, Apache-2.0 code), Qwen3-VL text encoder (LICENSE-Qwen.txt), LMS Sharp LoRA by Alissonerdx (its own terms). Open-H3-Regenerate-2K-IC is not an official MiniMax product. Powered by MiniMax H3.