Nodes/Minimax H3 Temporal Window & Spatial Tiles/Minimax H3 V2V Temporal Window & Spatial Tiles (mo-akkakk)
ComfyUI Node
Minimax H3 V2V Temporal Window & Spatial Tiles (mo-akkakk)
Put after the model, and route the CLIP through it into Reference to Video. Samples the video as overlapping temporal windows, one after another: each window is a normal H3 run on its slice of the plate, starts from the previous window's finished overlap frames (frame-0 guide), and the overlap is crossfaded. Optional spatial tiles, each following its crop of the plate. Works with KSampler and SamplerCustomAdvanced.
Minimax H3 V2V Temporal Window & Spatial Tiles (mo-akkakk)
- model
- clip
- plate
- vae
- MODEL
- CLIP
◄temporal_windowstrue►
◄window_frames124►
◄overlap_frames22►
◄short_final_windowfalse►
◄min_final_window_frames22►
◄carry_overlaptrue►
◄spatial_tilesfalse►
◄tile_width960►
◄tile_height544►
◄tile_overlap256►
◄crop_ref_images_per_tiletrue►
◄tile_audiogenerate_once►
◄color_matchfalse►
◄tile_carryown_tile►
◄plate_scale1.00►
Categorymodel/patch/minimax
Inputs (19)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| temporal_windows | BOOLEAN | true | Split the shot into overlapping temporal windows. Off = the whole length runs as one native H3 pass. |
| window_frames | INT | 12422–3600 | Frames per window (17k+5). Keep it inside what works natively for you, e.g. <= 192. |
| overlap_frames | INT | 225–1000 | Frames shared by neighbouring windows (5, 22, 39...). Carried into the next window as its frame-0 guide, then crossfaded. |
| short_final_window | BOOLEAN | false | Let the last window be shorter than window_frames instead of pulling it back to full length. Saves the frames that would be generated twice at the end, which matters most with large windows. Off = previous behaviour. |
| min_final_window_frames | INT | 225–3600 | With short_final_window on: the last window is never shorter than this (17k+5: 5, 22, 39, 56...). If the natural tail is shorter, the window is pulled back only far enough to reach it. |
| carry_overlap | BOOLEAN | true | Give each window the previous window's finished overlap frames (video and audio) as its frame-0 guide. |
| spatial_tiles | BOOLEAN | false | Split each window into overlapping tiles, for stage 1 or stage 2. Each tile follows its crop of the plate. Without a reference video, a full-denoise pass in tiles repeats the subject. |
| tile_width | INT | 960256–16384 | — |
| tile_height | INT | 544256–16384 | — |
| tile_overlap | INT | 2560–8192 | — |
| crop_ref_images_per_tile | BOOLEAN | true | With tiles, crop <Picture N> to each tile too. On when the picture is a frame of this shot; off for character/style references. |
| tile_audio | COMBO | generate_once | With tiles: generate_once = use the soundtrack generated by the first tile (instead of averaging one per tile). keep_input = output the incoming audio latent untouched (stage 2, when stage 1's audio is final). |
| color_match | BOOLEAN | false | Remove a small colour/brightness offset (per-channel median, clamped) between a tile and the tiles already blended into its overlap, and between a window and the previous one over their overlap. Off = no correction. |
| tile_carry | COMBO | own_tile | With tiles and carry_overlap: own_tile = each tile continues from its own unblended frames of the previous window, so blending errors are not fed forward (less drift over long shots). blended = continue from the blended picture (1.0.6 and earlier). |
| plate_scale | FLOAT | 1.000.25–1 | Only with the plate input connected. Size of each tile's plate piece relative to the tile. 1.0 = full detail (slowest: about 4x the reference tokens of no plate). 0.5 = a quarter of the reference tokens, much faster, still exactly aligned and sharper than without a plate. |
| clipopt | CLIP | Route the CLIP through here into Reference to Video so each window re-encodes the prompt with only its own slice of <Video N> (timestamps from 0), like a manual batch. Strongly recommended. | |
| plateopt | IMAGE | Optional, for tiles: the full-resolution version of the video you feed as ref_video_0 (same frames, same order, at least as long). Each tile's plate piece is then cut from it exactly and at full detail, instead of from Reference to Video's 768p copy. Needs vae. Feed it at about your output resolution (a 4K plate works but uses a lot of RAM). See plate_scale. | |
| vaeopt | VAE | Video VAE, used with plate to encode each tile's plate piece. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |
| CLIP | CLIP | — |