← sources

In Minimax H3 Reference Mode: If your model not following prompt - copy/paste prompt several times

source: reddit (reddit:comfyui)published 2026-10-09 02:25ingested 2026-10-10 02:37

retained text

I tried dozens of different ways of rephrasing prompt, but nothing made more effect than copy/pasting entire text. You can use this one for switching a character in a video, but you might need to add extra detail.

    subject_definitions:
    
    <Subject 1> is the girl in <Picture 1>.
    <Subject 2> is the girl in <Video 1>.
    <Video 1> is the movement reference for characters and reference for composition of [Shot 1].
    <Audio 1> is audio that plays in target video.
    
    summary:
    [reference generation] The target video shows <Subject 1> replacing <Subject 2> in <Video 1>, exactly copying motion pattern of <Subject 2>. Movement and composition perfectly synced with <Video 1> movement and composition.
    
    retention_analysis:
    <Subject 1> (appears in [Shot 1]): fully_preserved - the girl's identity and body shape are retained.
    <Subject 2>: reference - it's movement pattern and pose guides movement pattern and pose of <Subject 1>.
    <Picture 1>: partially_preserved - only reference to girl's identity, body proportions and appearance.
    <Video 1> (movement reference for [shot 1]): reference - movement pattern from video guides movement of <Subject 1> and <Subject 2>.
    <Audio 1>: fully_copy - plays in target video.
    
    detailed_description:
    The scene takes place in real world room. <Subject 1>'s pose and motion pattern is exact copy of pose and motion pattern of <Subject 2>.
    [Shot 1] <Subject 1>, replacing <Subject 2> in <Video 1>, dancing.
    
    overall_soundscape:
    N/A
    
    non_diegetic_music:
    N/A

Repeating prompt 10 times didn't help result more than 3 times.   
More detail - better: confirming camera position, character position, rotation and location. (ex. closeup shot of girl facing camera)(ex. replace "dancing" with "walking left")  
I'm using Res\_multistep sampler with sgm\_uniform scheduler: one of the fastest combinations without crazy hallucinations.   
I'm running "int4" version on Rtx 5070 with these starting arguments:  
\--enable-manager --disable-comfy-compiler --disable-pinned-memory  
  
Step increase and resolution increase didn't help overall adherence.

document fetched 2026-10-10 00:41 · canonical https://reddit.com/r/comfyui/comments/1x19om2/in_minimax_h3_reference_mode_if_your_model_not