comfyui-h3_temporal_sampler

H3 Temporal Sampler

ComfyUI custom nodes for seamless high-resolution MiniMax H3 video generation. Split an H3 audio-video latent that exceeds your VRAM into overlapping segments, denoise them consistently, and cache latents between runs.

This node is optimized for increasing output size (up to 2K), but it can also be used for longer videos — working around VRAM limits at the cost of longer render times.

Nodes

All nodes live under the h3 temporal sampler category.

H3 Temporal Sampler

Segmented SamplerCustomAdvanced for H3 AV latents. Splits the latent into overlapping time segments (grid-snapped to multiples of 5 latent frames) so each fits in VRAM, denoises each segment, and joins them seamlessly.

Inputs

| Widget | Type | Default | Notes | |---|---|---|---| | enable | BOOLEAN | True | OFF = one plain pass over the full latent, like SamplerCustomAdvanced. | | num_segments | INT | 4 | 1–10. How many overlapping segments to split the latent into — more segments fit larger sizes in VRAM. | | smart_bounds | BOOLEAN | False | ON = place boundaries at motion valleys (low-motion points) instead of even spacing. Falls back to even spacing if infeasible. | | overlap_frames | INT | 5 | 5–20, snapped to multiples of 5. Overlap between neighbours. Larger = smoother joins, more compute. | | blend_mode | COMBO | multiband | Join style for the final assembly: linear / adaptive / multiband (fine detail blended narrowly — best for fine textures). | | seam_lock | BOOLEAN | True | ON = freeze the leading frames of each segment (after the first) from the merged timeline via the denoise mask, so joins stay invisible. | | seam_lock_frames | INT | 2 | 1–4. How many leading frames to freeze; the freeze tapers off gradually (masks 0.0, 0.25, 0.5, 0.75). |

Outputs: output (LATENT), denoised_output (LATENT)

How it works

Each segment is sampled independently (one model init, shared across segments via equal-shape padding), then joined with the selected blend_mode. With Seam Lock on, the leading frames of each segment are frozen from the merged timeline via the denoise mask before sampling, so joins stay invisible. A per-boundary seam-quality report is printed (lower = cleaner; CHECK flags scores ≥ 1.0).

Notes

• Each segment is capped by a VRAM budget (validated reference: 34 latent frames at 84×144); if segments get too large, raise num_segments.
• Overlap is reserved up front, so padded segments never exceed the cap.

Tuning segments for your hardware

num_segments is your main lever: more segments mean shorter segments, and shorter segments use less VRAM — which is also how you push output size higher than a single pass could handle. If a render runs out of memory, raise num_segments; if every segment is comfortably small, lower it for a faster render.

The trade-off is time: each extra segment adds its own denoising pass (plus overlap work), so total render time grows with the segment count. As a rule of thumb, use the fewest segments that still fit — that is the shortest render your GPU can handle.

There is no universal number: the right count depends on your GPU's VRAM, the target size and the clip duration, so tune it yourself until you get a feel for what your card handles. The current defaults are tuned for the author's own system.

Demo

https://github.com/user-attachments/assets/955a1ff6-3759-4d37-a455-b27ec17e182e

Sample render — Stage 1 was short enough for a single pass, no Temporal Sampler needed. Stage 2 crashed on VRAM, so it was split into 4 segments (overlap 5, Seam Lock Frames 2) to get through. Segment cuts at approx. 2.1s, 4.9s and 7.7s of the 11.5s clip. Fine textures stay perfectly stable across all 4 segments, with no visible seams. Seam-quality scores: 0.67 / 0.36 / 0.15 (lower = cleaner).

Texture check across the segment boundaries: facial skin tone and lighting stay continuous, the tie's dot pattern keeps its size, spacing and alignment, and the jacket's houndstooth weave shows no breaks — only natural singing motion.

H3 Latent Cache (Save/Load)

Save and load H3 nested latents. ComfyUI's core Save/Load Latent nodes crash on H3's NestedTensor AV latent; this node round-trips the latent dict with torch.save / torch.load, preserving the structure exactly.

Inputs

| Widget | Type | Default | Notes | |---|---|---|---| | latent | LATENT | — | Live latent. Saved to a new file in SAVE mode; ignored in LOAD mode. | | load_mode | BOOLEAN | False | OFF = save the input to a new auto-numbered file and pass it through. ON = ignore the input, load the file chosen below. | | filename_prefix | STRING | h3_stage2_latent | Prefix for newly saved files (auto-numbered _00001, _00002, …). | | latent_file | COMBO | — | Saved file to load when load_mode is ON. |

Output: latent (LATENT)

Files are written as .h3latent.pt into ComfyUI's output directory. After saving new files, re-add (or duplicate) the node to refresh the dropdown list.

Installation

Option 1: Git clone (recommended)

cd ComfyUI/custom_nodes
git clone https://github.com/ehsrzv/comfyui-h3_temporal_sampler.git

Then restart ComfyUI.

Option 2: Download zip

1. Download the zip from the releases page (or the latest zip).
2. Extract it into ComfyUI/custom_nodes/ so you get: ComfyUI/custom_nodes/comfyui-h3_temporal_sampler/
3. Restart ComfyUI.

Option 3: ComfyUI Manager

Search for h3-temporal-sampler in ComfyUI Manager and install.

---

After installation, find the nodes under the h3 temporal sampler category. No extra dependencies — only ComfyUI itself and a MiniMax H3 setup.

Requirements

• ComfyUI (recent)
• MiniMax H3 model + a compatible guider/sampler setup
• The latent passed to H3 Temporal Sampler must be an H3 AV latent (nested video + audio samples)

Acknowledgments

Example workflow based on the workflow templates from LBH-123-AI/Minimax_h3_latent_Upscaler (Apache-2.0).

License

Apache 2.0