sandpies/ComfyUI-H3-Video-Upsampler

H3 Video Upsampler + h3upscale x2 LoRA

Code: GitHub. The LoRA is in lora/.

Upscale a finished MiniMax H3 clip (or any clip) in ComfyUI: the H3 latent upscaler enlarges it, then one light re-draw step in tiles restores detail. With the h3upscale x2 LoRA that step also gets the clip itself, at half the output size, as a guide -- so it restores the clip's own detail instead of inventing texture. Mouths, motion and the soundtrack come from your clip.

Install

1. Clone into ComfyUI/custom_nodes and restart ComfyUI.
2. Put the latent upscaler minimax_h3_latent_upscaler_3d_bf16.safetensors in ComfyUI/models/latent_upscale_models (LBH-123-AI).
3. Put h3upscale_v2_x2.safetensors in ComfyUI/models/loras (from Hugging Face).
4. Load example_workflows/H3_Video_Upsampler.json, pick a clip, queue.

Needs a ComfyUI with MiniMax H3 support and the H3 models (DiT, text encoder, video and audio VAE).

The node: H3 Video Upsampler

widget	what it does
megapixels	output size, aspect kept (2 = about 1920x1080)
denoise	how hard the one re-draw step works; 0.1 is the tested setting, 0 only enlarges
lora	h3upscale_v2_x2.safetensors = guided re-draw (detected from its header). Any other LoRA just sits on the re-draw
tile_tokens	VRAM. Largest sequence per tile; lower = more, smaller tiles, slower, less memory. 70000 is the default; try 40000 or 25000 on smaller cards
prompt	defaults to the caption the LoRA was trained on; keep it with that LoRA

• Wire the plain diffusion model loader as model when using the LoRA: it was trained on the bare DiT, and extra LoRAs under it re-add their own texture.
• audio_vae wired: the re-draw hears the soundtrack (best at 24 fps). Unwired: it hears silence. Either way you get your original soundtrack back.
• Any frame size and length: frames are fitted to 32 px and padded to a valid H3 length, then the padding is trimmed off again.

What the LoRA does and does not do

Tested by eye on H3 renders: guided at denoise 0.1, the 2 MP result beat the native render in a blind comparison. It is an upscaler: generating a video from noise with it flickers every 17 frames (it was trained on 5-frame clips); refining an existing clip, as this node does, does not.

Credits

Built in ComfyUI-Hand-Tie-Clips, where the same upsampler runs over a whole chain. Latent upscaler network: LBH-123-AI. Upscale recipe: bbaudio-2025's MMH3 Ultimate Upscale. Guide layout and dataset: Alissonerdx. See THIRD_PARTY_NOTICES.md.