ComfyUI Node
H3 Tiled Sampler Custom Advanced
Refine a long or upscaled MiniMax H3 clip with KSampler's settings, in overlapping tiles across the frame and windows along the clip that each fit the card, blended at every step into one clip. It does no upscaling itself. Each time window stays inside one scene, and the panel shows where the tiles sit. Requires a dict of video (or single image) + audio, and returns a dict.
H3 Tiled Sampler Custom Advanced
- model
- sigmas
- sampler
- conditioning
- latent
- latent
- denoised latent
◄seed0►
◄start_at_step0►
◄end_at_step10000►
◄tile_width576►
◄tile_height832►
◄overlap128►
◄window_frames0►
◄window_overlap17►
◄preview_methodtaeh3►
CategoryComfyUI-GTNodes/samplers
Inputs (14)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | The MiniMax H3 model, after any LoRA. | |
| sigmas | SIGMAS | — | |
| seed | INT | 00–18446744073709550000 | The noise seed, `0` as good as any. |
| sampler | SAMPLER | — | |
| conditioning | CONDITIONING | The clip's positive conditioning. | |
| latent | LATENT | The H3 dict with upscaled vimage latent to refine + audio | |
| start_at_step | INT | 00–10000 | — |
| end_at_step | INT | 100000–10000 | — |
| tile_width | INT | 57664–16384 | Pixels a tile spans across the frame, as `576`. |
| tile_height | INT | 83264–16384 | Pixels a tile spans down the frame, as `832`. |
| overlap | INT | 12832–4096 | Pixels neighbouring tiles share, as `128`. |
| window_frames | INT | 00–10000 | Frames a time window spans, as `141`, or `0` for the whole clip or a single image. |
| window_overlap | INT | 171–1000 | Frames neighbouring time windows share, as `17` for a video or `1` for a single image. |
| preview_method | COMBO | taeh3 | "taeh3" requires "taeh3.safetensors" in models/vae_approx, from kijai, and sets "max_preview_resolution" at 1024. The animated preview requires https://github.com/kosinkadink/ComfyUI-VideoHelperSuite, then: Settings -> 🎥🅥🅗🅢 -> "Display animated previews when sampling" ON. You can also change video preview's fps here. For a single image VHS is not required. "taeh3" previews have been tested both with single image and video generation. Decode tests (on a rtx 3060 12gb): - 8 seconds, 448x704: the sampler stopped each step for 1.3 seconds; - 8 seconds, 576x832: the sampler stopped each step for 2.1 seconds |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| latent | LATENT | The H3 dict with refined latent image + audio |
| denoised latent | LATENT | The denoised H3 dict (if the sampling was stopped before the last sigma) with refined latent image + audio |