ComfyUI Node
VRGDG H3 Chunked Pass 2
MiniMax H3 second pass in temporal windows: the normal learned upscale and SamplerCustomAdvanced for one window at a time, assembled on the CPU, with the first-pass audio preserved exactly. The Video Builder uses it for scenes longer than its Pass 2 chunking threshold.
VRGDG H3 Chunked Pass 2
- latent
- model
- conditioning
- upscale_model
- sampler
- sigmas
- head_latent
- head_vae
- refined_av_latent
- chunk_report
◄scale2.0►
◄chunk_seconds3.0►
◄overlap_seconds1.6►
◄overlap_mode▾►
◄seed42►
◄tail_trim_seconds0.7►
◄max_window_tokens0►
◄upscale▾►
CategoryVRGDG/MiniMax H3
Inputs (16)
| Name | Type | Default | Description |
|---|---|---|---|
| latent | LATENT | The first-pass joint video+audio latent. | |
| model | MODEL | — | |
| conditioning | CONDITIONING | — | |
| upscale_model | VRGDG_MINIMAX_H3_LATENT_UPSCALE_MODEL | — | |
| sampler | SAMPLER | — | |
| sigmas | SIGMAS | — | |
| scale | FLOAT | 2.01–4 | — |
| chunk_seconds | FLOAT | 3.00.1–120 | — |
| overlap_seconds | FLOAT | 1.60–30 | — |
| overlap_mode | COMBO | locked pins the previous refined overlap while the next window samples. crossfade refines both versions of the overlap and blends them. | |
| seed | INT | 420–18446744073709550000 | — |
| tail_trim_secondsopt | FLOAT | 0.70–10 | Throw away this much of each window's tail (snapped to 0.708 s groups) and let the next window redo it with refined history on the left and content on the right. Needs an overlap of at least two groups (1.6 s). 0 keeps every window's tail. |
| max_window_tokensopt | INT | 00–2000000 | Size windows by video tokens instead of seconds. 0 turns the budget off. A budget below the smallest window (7 latent frames) uses that window and says so. |
| upscaleopt | COMBO | per_window upscales only the current window, so upscaler memory is one window. whole_clip upscales the entire clip first: window edges get real neighbours, but upscaler memory grows with the clip. | |
| head_latentopt | LATENT | Masked continuation: the predecessor window from VRGDG H3 Load Latent (masked_av). It becomes the locked first tokens of the refined clip. | |
| head_vaeopt | VAE | Video VAE, used only to resize a head that was saved at another size. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| refined_av_latent | LATENT | — |
| chunk_report | STRING | — |