IronLeo78-MTV-Prompt-Generator

9 pictures + 1 video + your prompt in one node. Pictures (the nine slots on one wire) goes to the MTV node's Pictures input; MTV_Prompt (written by the LLM from the pictures, the video, the Preset and your prompt) goes to the MTV prompt.

By ironleo78-sys·Created 6 days ago·Updated about 4 hours ago· 3

Pictures
MTV_Prompt
seed

◄Modev2v►

◄Song modefalse►

◄image_1►

◄image_2►

◄image_3►

◄image_4►

◄image_5►

◄image_6►

◄image_7►

◄image_8►

◄image_9►

◄video►

◄Start_From0.0►

◄Duration5.0►

◄FPS_To_LLM2 FPS►

◄Video_Size512x288►

◄PresetNone►

◄prompt►

◄LLMHuihui-Qwen3-VL-4B-Instruct-abliterated-int8_convrot.safetensors►

◄Use LLMtrue►

◄Keep LLM in RAMtrue►

◄Momentum rulestrue►

◄Seed0►

◄song_facts—►

CategoryIronLeo78/MTV

Inputs (24)

Name	Type	Default	Description
Mode	COMBO	v2v	v2v: the presets work on the video you load (Picture 1 = the first character, as before). i2v: no video; Picture 1 becomes the SCENE that replaces Video 1 and the characters move up one slot (preset 1: Picture 2 = the new character). A loaded video is ignored in i2v.
Song mode	BOOLEAN	false	ON: the song is the clip's sound, put straight into the video model (the MTV's Choose Audio latent is set to Custom Audio and Next Segment On Run to ON). The LLM writes ONE plain prompt (guide/Song.md, not the MiniMax six sections) from the preset's song version: every singer lip-syncs exactly, every musician plays exactly in time, and the story can follow the song. The Song presets switch it ON.
image_1	STRING		—
image_2	STRING		—
image_3	STRING		—
image_4	STRING		—
image_5	STRING		—
image_6	STRING		—
image_7	STRING		—
image_8	STRING		—
image_9	STRING		—
video	STRING		—
Start_From	FLOAT	0.00–36000	Start of the video window, in seconds.
Duration	FLOAT	5.01–20	Length of the video window, 1-20 seconds.
FPS_To_LLM	COMBO	2 FPS	How many frames per second of the video window go to the LLM (the video is read at 24 fps and every Nth frame is kept: 6 FPS = every 4th, 1 FPS = every 24th). More FPS = a more exact timeline, a slower prompt. At most 120 frames go to the LLM; the info line says when a window is above that.
Video_Size	COMBO	512x288	3 options: 512x288, 672x384, 896x512
Preset	COMBO	None	Scenario preset for the LLM prompt generator (files in the pack's presets/ folder). None = no preset. The preset travels inside the LLM_Frame_Info output.
prompt	STRING		Your short request, e.g. 'Replace the main character of Video 1.' The preset and the frame information are added for you.
LLM	COMBO	Huihui-Qwen3-VL-4B-Instruct-abliterated-int8_convrot.safetensors	The vision LLM (a Qwen3-VL text encoder in models/text_encoders).
Use LLM	BOOLEAN	true	ON: the LLM writes MTV_Prompt from the pictures, the video, the Preset and your prompt. OFF: the LLM is not loaded or run, and MTV_Prompt is your prompt exactly as typed (Pictures still go to the MTV).
Keep LLM in RAM	BOOLEAN	true	OFF: the LLM is unloaded after each run and read from disk again next time. ON (default): it stays in RAM (about the size of its file, several GB) and only leaves the GPU, so the next run starts faster. Choosing another LLM file, or switching this OFF, gives the RAM back.
Momentum rules	BOOLEAN	true	ON: when a preset continues Video 1, the prompt must restate the exact last-frame state (pose, camera, direction and speed of the motion) and carry that momentum on. Ignored by presets that do not continue a video.
Seed	INT	00–18446744073709550000	Seed of the LLM. Same seed + same inputs = the same MTV_Prompt. Change it, or set control after generate to randomize, to get another text.
song_factsopt	STRING		v0.12.0: optional. IronLeo78 Undercover's facts output: the LLM is told which part of the song this clip covers (verse / chorus ...) and the lines sung in it (estimated from the lyrics).

Outputs (3)

Name	Type	Description
Pictures	IRONLEO78_PICTURES	—
MTV_Prompt	STRING	—
seed	INT	—