MiniMax H3 Plan Transitions

The difference between an H3 video that feels like a film and one that feels like a slideshow is almost entirely in the transitions. Whether scene 3 carries on the shot, cuts hard, or cuts while keeping the sound changes how the thing reads - and picking eight of those by hand, twice a night, is exactly the kind of job people stop doing properly. This node reads your scenes with a language model and picks them for you.

How it works

You hand it the scene prompts in order, plus a vlm_clip - a language model loaded through ComfyUI's own CLIP path - and it answers with one decision per scene from the second onward: a cut, a carry of the same shot, a cut that keeps the sound, or a cut that keeps the cast, with or without the sound. The choices map onto the same transition vocabulary MiniMax H3 Conditioning's rows use (carry, handoff, reference (video), reference (sample), cut, and the audio-carrying variants), and each entry also says why, which is genuinely useful when one of them is wrong.

It's a text node. Nothing here samples video and no model is loaded by a render - this decides, the loop renders. The seed picks which draw of the model's choices you get; keep it and you get the same plan back, which is how you compare two prompts without also comparing two sets of transitions.

If you're using the Prompt Timeline window, you can skip the wiring entirely: Plan transitions in the timeline bar runs this node and writes each row's transition and overlap for you.

The inputs that matter

• vlm_clip - from Load CLIP. qwen3vl_8b_fp8_scaled.safetensors or qwen_3_8b_fp8mixed.safetensors are the names the node suggests; put them in models/text_encoders.
• scenes - the prompts in order, divided by a line of dashes exactly as MiniMax H3 Scene Writer's prompts output produces them. You can also paste the Scene Writer's scenes JSON.
• seed - which draw. Same seed and inputs, same plan.
• The generation dials: temperature (0.3 by default - low, because you want the likeliest reading rather than a creative one), top_k, top_p, min_p, repetition_penalty, thinking for reasoning models such as Qwen3, and fast_decode.

Two outputs. transitions is JSON, one entry per scene from the second, with the scene number, the transition, the continuity and overlap it implies, and the reason. report is a count of how many of each were chosen - a quick sanity check that you didn't get fourteen hard cuts.

Installing it

It ships in WAS Node Suite v3: ComfyUI Manager, search WAS Node Suite v3, or:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui.git

ComfyUI 0.14.0+ and Python 3.10+, and the pack installs nothing - no llama.cpp, no server, no separate Ollama process. The model is a ComfyUI CLIP, loaded by core Load CLIP.

That does mean real VRAM is in play. A Qwen3-VL 8B in fp8 is a few gigabytes resident, which is the standard local-LLM-in-ComfyUI arithmetic: you're budgeting for a second model beside the video model. If you're tight, a smaller quant is the usual answer, and the pack's fast_decode option keeps the decode on the fixed-cache path where the model allows it.

Where it goes wrong

Garbage or refusals in the output. Turn temperature down before anything else. 0.3 is the default for a reason; the node wants a deterministic picker, not a novelist.

Every transition is cut. Either the model can't see continuity to preserve, or the scenes share none. If scene 4 is meant to continue scene 3's shot, its prompt has to say so - the planner reads prompts, not your intentions. References and keyframes declared in your asset chain are part of what it has to work with, so wire the MiniMax H3 Asset chain in before planning.

"Nothing happened" after execution. Correct - it produced text. Look at transitions (use a preview or a show-text node) or at the rows in the Prompt Timeline, not at the canvas.

thinking: true and it takes forever. Reasoning models think before answering, and that's tokens on your card. It's a good switch for a two-scene plan and a bad one for a twenty-four-scene plan.

The format matters. Dashes between prompts, or the Scene Writer's JSON - not a numbered list you typed by hand. Feed it the wrong shape and it plans transitions for whatever it thinks the scenes are.