Home Creators Posts Import Register Favorites Logout

Downloads

Content

Video : https://youtu.be/Y1m6R8shJcw

Hey everyone.

New video is up, and this one is for anyone who has fought an image model over composition. You write the prompt, you hit generate, and the subject lands dead center, standing like a robot.

Over the weekend I built two custom ComfyUI nodes for MiniMax H3 Image. One is a canvas where you draw placement boxes over your layout, this character here, that character there. The other turns your boxes into a real layout prompt for the model. Each box gets a short description of who is inside it and what they are doing, and a local 9B language model assembles the final prompt for you.

The finding that surprised me: raw coordinates alone give you stiff results. Once relational language is mixed in with the x and y positions, characters blend into the background far better. My demo scene, two people arguing in a hotel hallway, reads like a proper start frame for a video scene.

This is experimental, and I say that openly in the video. Some scenes still miss, and you will be resizing boxes by hand, which is common sense as a human user, right? The nodes are open on GitHub if you want to try them and break them.

https://github.com/benjiyaya/ComfyUI-H3Image-LayoutKit/

See you in the next one.

— Benji

Files

MiniMax H3 Image Bounding Box Layout Custom Node in ComfyUI

I built two ComfyUI custom nodes for MiniMax H3 Image that let you place subjects with bounding boxes: an H3 Bounding Box Canvas for drawing placements interactively, and an H3 Layout Prompt Builder that turns your boxes into a model-ready layout prompt. Instead of fighting an all-text prompt, you draw a box for each character or reference image, write a short description per box, and the nodes convert the x and y coordinates (0-1000 grid) plus relational language into the final prompt. The coordinate mode can run relational, coords only, or both. Both mode blends subjects into your background reference far better than raw coordinates alone, which tend to produce stiff, centered, robot-like placements. I walk through the full workflow, generate a two-character hallway scene from reference sheets, adjust box sizes, and explain why I deliberately break the official MiniMax H3 prompt-writing rules for this kind of task. This is for ComfyUI users who generate multi-subject images from reference sheets, especially anyone preparing clean start frames for video generation. No training needed, just the nodes and a local 9B language model for the prompt generation step, following the same pattern as the ComfyUI docs with Qwen 3.5 9B. Bounding box control is a concept that works in the video models through ControlNet-style conditions and colored placement boxes. Here I bring the same idea into an image model while it is still experimental. Some scenes render perfectly, some need you to resize your boxes like a sensible human, and a LoRA may be the next step to tighten adherence. The nodes, prompt template, and guard conditions are all open on GitHub so you can try, break, and extend them yourself. Resources: - GitHub nodes: https://github.com/benjiyaya/ComfyUI-H3Image-LayoutKit/ Workflow for freebies: https://www.patreon.com/aifuturetech/posts/minimax-h3-image-175485569 Calliope - Local-first Agentic Idea-to-video Studio https://github.com/benjiyaya/Calliope If You Like tutorial like this, You Can Support Our Work In Patreon: https://www.patreon.com/c/aifuturetech #MiniMaxH3 #ComfyUI #AIImageEditing #BoundingBox #ComfyUINodes

Comments

No comments found for this post.