arkennemasis Hailuo Scene (one clip, start to finish)

ArkHailuoScene renders ONE scene end to end with Hailuo MiniMax H3 - condition, sample, decode video and audio, mux and write the file - then frees everything. It's the "one scene at a time" node that makes the arkennemasis video chain's per-scene loop hold memory at a single scene instead of accumulating every scene's latents until the graph runs out of VRAM.

It wraps the full H3 pipeline into a single node: video VAE, audio VAE, sampler, sigmas, and the scene's first-frame still go in, a finished VIDEO file comes out. That "then free everything" is the part people underrate - inside a loop that renders scene after scene, it's the difference between a 12-scene film fitting in 16GB and dying at scene four.

The inputs that matter

The model-side inputs are model, clip, vae (the video VAE), audio_vae (the audio VAE - the thing that gives the clip its soundtrack; H3 generates audio natively, which is why there are two VAEs), sampler, and sigmas - wire them from the pack's Video Model setup or any standard H3 nodes.

The scene-side inputs are where the workflow lives:

• image - this scene's still, the first frame of the clip.
• prompt - the scene's text.
• width / height - 1280×736 by default.
• length - frames for this scene, wired from Scene List so each clip runs its own duration. It's forceInput on purpose: a widget carrying a link occupies a widgets_values slot, and slot-count disagreement shifts every later value in saved workflows. (This pack is religious about not corrupting saved graphs, and it shows.)
• seed - a base seed, mixed with this scene's prompt, so every scene of a run samples differently instead of all sharing one noise pattern.
• reseed_each_run - default on: draw fresh noise every run. Off reproduces the same clip down to the audio - including the same garbled speech, which the tooltip warns "looks like the scene is locked." Turn it off only to reproduce a specific take exactly.
• filename_prefix - relative to ComfyUI's output dir; each call appends the next counter so clips land in scene order.
• last_frame - optional closing frame. Give it one and H3 animates the transition from the first frame to this one instead of inventing motion from a single still. That's the first-frame-to-last-frame transformation.

Outputs: video (the VIDEO type) and path.

A licensing note worth knowing

H3's open weights ship under the MiniMax H3 Community License, which geofences out the US, EU, UK and South Korea - users in those regions are not licensed to run the local weights at all (the hosted Hailuo API remains globally available). That's a real constraint on who can legally use this node, and it's the kind of thing worth checking before you build a production pipeline on it. The pack doesn't hide it, and neither should you.

Installing

It's part of comfyui-arkennemasis:

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis
pip install replicate httpx

restart, under arkennemasis/Video. You'll also need the H3 model files and ComfyUI's own MiniMax H3 support in place - this node orchestrates them rather than shipping the weights.

Where people get burned: forgetting that length is a frames count on H3's 17k+5 frame grid (Narration Length exists precisely to compute this from a voice-over), and expecting the single-still mode to produce motion it wasn't asked for - no last_frame, and H3 invents the motion from the still, which is a coin flip you should know you're taking.