Comfyui-MMH3-UltimateUpscale

Comfyui-MMH3-UltimateUpscale

Upscale long, high-resolution MiniMax H3 video on a VRAM-limited GPU — in a single node.

This node re-samples (enhances / upscales) an already-denoised MiniMax H3 AV latent through the full auto pipeline under tight VRAM constraints: it processes the clip with temporal chunking (so arbitrarily long videos fit in memory) + spatial tiling (so arbitrarily high resolutions fit in memory), keeping peak VRAM bounded to a single tile, while preserving the audio track intact.

MiniMax H3 generates video as a nested latent that bundles 24-channel video and 32-channel audio in one tensor. Standard ComfyUI upscale nodes do not understand this structure. MMH3 Ultimate Upscale wraps the entire temporal split -> latent upscale -> spatial split -> per-tile sampling -> spatial stitch -> temporal stitch loop into one node, so you can upscale a finished H3 clip the same way you would upscale a normal latent — without breaking audio and without running out of VRAM even on small cards.

---

Changelog

• 20260923 - Add new node MMH3 Tiled Diffusion (Experimental). It can be used to replace spatial split, basically without any seam defects.
• 20260905 - Latent upscale checkpoints are now .safetensors only. Pickle formats (.pth / .pt / .ckpt) are no longer listed or loaded, because torch.load() on a pickle checkpoint can execute arbitrary code and weights_only=True is not a reliable mitigation on torch < 2.6 (CVE-2025-32434). If you were using minimax_h3_latent_upscaler_3d_fp32.pth, convert it once (see Upscale model format) or switch to the fp16 / bf16 .safetensors build. The scanner now also searches every registered latent_upscale_models directory (including extra_model_paths.yaml) instead of only the first one.
• 20260831 - Rearrange the code structure. Bugs fix and add some features. Add dynamic fade to improve seam removal when denoise step>1. Add brightness match to to help reduce brightness flickering.
• 20260829 - Add new nodes about Fun ControlNet to enhance performance. Two nodes MMH3 Fun Controlnet Inpaint and MMH3 Spatial Inpaint Params, which help seamless spatial stitching and better consistance in higher denoise condition. To use these nodes you have to merge kijai's pull request in comfyui.
• 20260825 - New experimental LTX25 Ultimate Upscale node. Built on top of the MMH3 pipeline, it now also supports LTX2.5 nested AV latents (video [B,128,T,H,W] + audio [B,C,time,freq]) in a single node: temporal split -> latent upscale (fixed 2x model upscale, then interpolated to the target width/height) -> spatial split -> per-tile sampling -> stitch. Three optional param nodes are provided: LTX25 Latent Upscale Params, LTX25 Temporal Split Params, and LTX25 Spatial Split Params. Audio is buggy so you should use original audio latent. These nodes are highly experimental, so don't rely on them.

---

Features

• Upscale long + high-res video on limited VRAM. The core design goal: temporal chunking keeps arbitrarily long clips in memory, spatial tiling keeps arbitrarily high resolutions in memory, and only one tile is sampled at a time — so peak VRAM stays at a single tile regardless of video length or output resolution.
• One node, full pipeline. Temporal chunking (outer loop), optional latent upscale, spatial tiling (inner loop), per-tile diffusion sampling, then spatial + temporal stitching — all driven by a single MMH3 Ultimate Upscale node.
• Temporal chunking for long videos. A long clip is cut into overlapping time chunks; each chunk is processed independently and stitched back together.
• Two upscale modes per chunk:
  • H3 3D model-based upscaler (MMH3 Latent Upscale with Model Params) — uses the minimax_h3_latent_upscaler_3d_*.safetensors chec