MiniMax H3 Trajectory Control Render / 轨迹控制视频 (Advanced)

The Plan node thinks in numbers; this one thinks in pixels. Give it a trajectory_plan and it renders an actual IMAGE batch - one frame per frame of your clip - that shows where every object's box is at every moment. That video is what you feed into H3 Fun Control, the ControlNet-style conditioning H3 ships with, and it's how a bunch of keyframes become a "this object drifts from the left edge to the right over two seconds" instruction the model can see.

Two honest caveats baked into the design: the Render node creates no audio and owns no H3 attention. It produces a conditioning video; the Fun Control Apply node (or your own canny/depth/pose preprocess) does the actual steering. It's a drawing tool, and it knows it.

How it works

It walks the plan's tracks and, for every frame, draws each object's interpolated bbox onto a canvas filled with background_level (0 = black). Three render_modes:

• reference_sprite (default) - pastes your actual reference images into each box, resized to fit and masked by an optional alpha, with per-keyframe strength controlling opacity. This is "the subject travels the path" - the sprite moves frame to frame. Requires reference_images connected or it errors immediately.
• soft_region - fills each box with a solid per-object color (a palette, one color per track), feathered by feather (a 0..0.2 sigmoid softness, 0 = hard edge). Good for "this region should contain the thing" conditioning.
• box_outline - like soft_region but hollow: the outline width is line_width (in canvas pixels). The lightest touch; the model just knows where the box is.

Optional reference_masks let you cut sprites to their true silhouette instead of a rectangle - one mask per object or a single shared one.

Outputs that matter:

• control_video - the IMAGE batch, one frame per plan frame. Wire it into MiniMaxH3FunControlApplyT8Advanced's control_video input (start with control_kind=custom), or preprocess it (canny/depth/pose) for the classic control route.
• trajectory_masks - a MASK batch of the union coverage; handy if you want to composite the path onto something else.
• preview - a 12-frame sample so you can eyeball the motion without previewing the whole clip.
• report_json - status, mode, frame count, plan hash, and the recommended connection.

Install

Same pack as everything else here: Manager → search "MiniMax H3 Audio T8" → install → restart, or

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

No pip installs - requirements.txt is intentionally empty; ComfyUI supplies the stack. To actually use the output you also need the H3 Fun Control model from Kijai/MiniMax-H3-experimental (recent weights in models/model_patches, older ones in models/controlnet - the pack accepts both contracts). And the usual requirement: a current ComfyUI, updated alongside the pack, or all T8 nodes show red.

Gotchas

• reference_sprite with no reference_images is an instant error - connect them first. One image shared by all objects, or one per object in order.
• Keep width/height multiples of 32 and the frame count on H3's 17n+5 grid (22, 39, 56…) or the Plan node upstream refuses to build.
• The README's general H3 advice applies: don't stack multiple attention/sampler-takeover nodes on top of Fun Control, and if you OOM, drop resolution/frames before blaming the conditioning.
• Realistically this is an EXP experiment. The sprite mode is the one that feels magical (an image physically walking your path); soft_region is the safest if H3 ignores the boxes. A/B a couple of modes on a fixed seed before you commit to a shot.