Skip to content

MiniMax-H3: feed reference images to the Qwen text encoder (semantic/timed <Picture N>) #1952

Description

@laurentvv

ComfyUI has a technique (ethanfel's ComfyUI-MiniMaxH3-Timed-References) where images are presented only to the Qwen text encoder as <Picture N> — semantic slots, no VAE encoding, optionally timestamped ("at 2.0 s, show the red car"), including frames picked from a video with their real PTS. It enables "shot-planning" prompts: a storyboard of timed intentions, distinct from the native VAE-encoded <Video 1> reference.

In sd-cli, as far as I can tell, the --llm Qwen3-VL GGUF only receives text today (images only flow through --ref-image/--ref-video -> VAE). Would you consider an optional --prompt-image <file>[:seconds] that injects images into the text-encoder pass as vision tokens with <Picture N> tags?

Open question on my side: does the qwen3vl_32b_minimax_h3 GGUF keep the vision projector usable (or is it text-only after quant/adaptation)? If mmproj weights are needed, I can test that path.

Use case: multi-shot H3 videos planned as a timed storyboard without burning native reference slots.

Environment: stable-diffusion.cpp master-841 win vulkan-x64, Windows 11, AMD RX 6950 XT 16 GB (RDNA2).

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions