MiniMax-H3 Multishot — chained shots, one master, with audio — v1.0 — v2.6.3

Both samplers explain the reference images. Attached photos arrive labelled , ... and only one of the two samplers ever told the model what those labels meant. Now both do - and a frame carried between chained shots is declared as the shot's starting frame instead of being mistaken for a face reference.

Keyframe modes send the documented alignment line. First frames, last frames and boundary plates now announce themselves the way the model was trained to read them. Three paths sent nothing before.

Removed a self-defeating negation. The reference block used to say a second character is "never blended with" the first. There is no negative prompt at cfg 1.0, so that sentence only put the idea of blending INTO the render. It now states what each person keeps instead.

Auto Refs: five characters no longer cast three

The automatic scan stopped at the first three characters it found. With five, two of them got no photos while the writer still pointed them at "the reference photographs" - and a character pointed at photos that do not exist renders as a random person. The scan now takes up to nine characters and splits the model's nine photo slots across everyone it matched (2 characters keep 3 photos each, 4 get 2, five or more get 1), and prints the split. Fewer characters per run still holds a face more firmly - but every named character now casts somebody.

One number for window length

In extend-take mode the frames_per_shot box was silently ignored - a window sized for your card took over. Now window = auto follows the frames_per_shot you typed (snapped to the legal grid, and the console says so), picking an explicit window still wins, and the old card-aware sizing is still there as fit this card (VRAM auto) - choose it when you want the fewest joins that will not thrash rather than an exact length. The shipped workflow is set to 243, which is exactly what it rendered before.

Remote encoding: fails fast, starts fast

A wrong or unedited address on the remote text encoder used to surface at the first encode - AFTER the LLM writer had spent minutes on a script that then went in the bin. The node now checks the address the moment it runs: empty box, untouched placeholder, or unreachable host each fail in about a second, with the fix named in the message.

First shot no longer crawls. With the encoder on another box, models left over from your PREVIOUS run were never cleared, the memory planner saw a full card, and shot 1 ran 2-3x slower than shot 2 (measured 65 s/it against 27). Both samplers now sweep leftovers once before the first model load and print what they cleared.

New in the pack

H3 Retake - redo one stretch of a finished clip and keep the rest. Load the clip, set a time window, write a prompt for that moment. Picture and sound are independent: redo both, keep the performance and change the picture, or keep the picture and change the line.

The prompt picker now lists EVERY registered prompt folder - a corpus added through extra_model_paths.yaml used to sit invisible.

Writer rules: recurring characters keep a readable face in frame and every shot ends on the face (identity re-locks from a shot's closing frames - a shot that ends on the back of a head hands the next shot a stranger); anything uncanny never turns to the lens and the world never reacts on cue; the camera is directed in the model's own trained vocabulary, with "holds a static shot" as the way to hold still - never "does not move", which freezes the frame.

Writer node: local Ollama endpoints get a proper context size - the 4096 default silently truncated long system prompts.

Text encoder · Unknown