Home Creators Posts Import Register Favorites Logout

Downloads

Content

Video : https://youtu.be/eiJbJMFU52M

Hey everyone.

New video is up, and this one is for anyone who has ever squinted at a soft MiniMax H3 clip and wished there was a proper way to finish it.

NVIDIA just open-sourced Open-H3-Regenerate-2K, an open version of the 2K regeneration step that only existed as a closed API until now. You feed it a low-res draft and it redraws the detail natively at 2K in three diffusion steps. In the video I run it in ComfyUI two different ways and put the outputs side by side.

What surprised me: NVIDIA's blinded test had people prefer it over the official API 87% of the time. And on my machine, my usual two-stage latent upscale with the extracted LoRA looked sharper than the full custom-node pipeline with the giant original file. Sometimes the small rig wins.

Faces in far shots are where you see it most. The 480p draft cannot even tell you who the character is. The 2K pass brings the face back, even mid-walk.

The free two-stage workflow from the video is attached, with a screenshot of the graph. Supporters, the full run guide with both methods and the compare workflow is in the supporter post.

- Benji

Resources:

- Project page: https://nvlabs.github.io/Sana/Open-H3-Regenerate-2K/

- NVIDIA GitHub (Sana, sol-engine): https://github.com/NVlabs/Sana/tree/sol-engine/models/Open-H3-Regenerate-2K

- Official model repo: https://huggingface.co/Efficient-Large-Model/Open-H3-Regenerate-2K

- Kijai extracted refiner LoRA rank 384 (~5GB): https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/loras/minimax_h3_open_regenerate_2k_refiner_rank_384_bf16.safetensors

- Kijai extracted refiner LoRA rank 16 (~155MB): https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/loras/minimax_h3_open_regenerate_2k_refiner_rank_16_bf16.safetensors

- ComfyUI-Open-H3-Regenerate custom nodes: https://github.com/rdh9616/ComfyUI-Open-H3-Regenerate

- MMH3 Ultimate Upscale pack (method 1): https://github.com/bbaudio-2025/Comfyui-MMH3-UltimateUpscale

- MiniMax H3 latent upscaler model: https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler

- LMS Sharp LoRA: https://huggingface.co/Alissonerdx/Minimax-H3-ComfyUI/blob/main/loras/minimax_h3_lms_v1.0_r64.safetensors

- MiniMax H3 VAE and text encoders: https://huggingface.co/Comfy-Org/MiniMax-H3

- H3 hybrid diffusion model (normal): https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models

Files

MiniMax H3 Video to 2K With Open-H3-Regenerate-2K In ComfyUI

NVIDIA Open-H3-Regenerate-2K in ComfyUI: turn a 480p MiniMax H3 draft into native 2K video in three steps, with two tested methods and every file linked. NVIDIA just open-sourced the regeneration step that the official MiniMax H3-Regenerate-2K API keeps closed. Open-H3-Regenerate-2K takes any low-resolution draft, 384p to 768p, and redraws the detail natively at 768p, 1080p, 2K or even 4K in a three-step diffusion pass. In this video I walk through the official Hugging Face repo, show what actually loads in ComfyUI, and build two working methods: method 1, my usual two-stage latent upscale with the MMH3 Ultimate Upscale pack and Kijai's extracted rank-384 refiner LoRA, and method 2, the ComfyUI-Open-H3-Regenerate custom node path that loads the original 12GB LoRA with the sharp_curve_branches file and the fast StreamRender H3 VAE decoder. This is for anyone running MiniMax H3 locally in ComfyUI who wants 2K output without the hosted API. You need a current ComfyUI with the native MiniMax H3 nodes, the two custom node packs linked below, and roughly 6GB of LoRA weights for the method I recommend. The setup sections cover the loaders, the SoL Attention setting that the NVIDIA engine needs, and where the LMS Sharp and Sharp Curve branches LoRAs sit in the chain. Why it matters: NVIDIA's own blinded user study preferred this open model over the official API in 87% of 200 judgments, and it runs a 5-second 2K refine in about 22 seconds on one B200. On my consumer GPU the three-step regenerate took about a minute per 2K clip, faster than my usual latent-upscale-only second stage. Honest finding from the tests: the simple two-stage method with the extracted rank-384 BF16 LoRA beat the full 12GB custom-node pipeline on sharpness, faces in far shots, hands and knuckles. The 12GB FP32 file is not worth the download for consumer GPUs. Resources: - Project page: https://nvlabs.github.io/Sana/Open-H3-Regenerate-2K/ - NVIDIA GitHub (Sana, sol-engine): https://github.com/NVlabs/Sana/tree/sol-engine/models/Open-H3-Regenerate-2K - Official model repo: https://huggingface.co/Efficient-Large-Model/Open-H3-Regenerate-2K - Kijai extracted refiner LoRA rank 384 (~5GB): https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/loras/minimax_h3_open_regenerate_2k_refiner_rank_384_bf16.safetensors - Kijai extracted refiner LoRA rank 16 (~155MB): https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/loras/minimax_h3_open_regenerate_2k_refiner_rank_16_bf16.safetensors - ComfyUI-Open-H3-Regenerate custom nodes: https://github.com/rdh9616/ComfyUI-Open-H3-Regenerate - MMH3 Ultimate Upscale pack (method 1): https://github.com/bbaudio-2025/Comfyui-MMH3-UltimateUpscale - MiniMax H3 latent upscaler model: https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler - LMS Sharp LoRA: https://huggingface.co/Alissonerdx/Minimax-H3-ComfyUI/blob/main/loras/minimax_h3_lms_v1.0_r64.safetensors - MiniMax H3 VAE and text encoders: https://huggingface.co/Comfy-Org/MiniMax-H3 - H3 hybrid diffusion model (normal): https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models - Free 2-stage workflow from this video: https://www.patreon.com/aifuturetech/posts/173747267?utm_source=youtube&utm_medium=video&utm_campaign=202610 Calliope - Local-first Agentic Idea-to-video Studio https://github.com/benjiyaya/Calliope If You Like tutorial like this, You Can Support Our Work In Patreon: https://www.patreon.com/c/aifuturetech

Comments

No comments found for this post.