minimaxH3DYTsc6.0 Director

One node, a whole studio

minimaxH3DYTsc6.0 Director (class MinimaxH3DYTScDirector) is the reason this pack exists. It is not a loader or a sampler - it's conditioning, sampling, decode, optional refinement and saving, all inside a single node with its own timeline UI drawn on the node body. You give it a prompt and a frame count and press run.

That's a fork in the road. If you like wiring graphs, the pack's MinimaxH3DYTScConditioning node plus ComfyUI's official H3 nodes does everything this does, visibly. The Director is for shots in a list - several prompts, several durations, reference material per shot, continuation from the previous segment's last frame - without 200 nodes of canvas.

The task_type enum has two entries, and they lead to genuinely different machinery. t2i is a text-to-image workspace that does not touch H3 at all: it runs a Krea2 Turbo pipeline with its own model card, and H3 checkpoints are ignored there. r2v is the video side, and despite the name it's a unified workspace - no reference material means plain text-to-video, keyframes mean i2v/fl2v, and reference images/videos/audio route through H3's reference conditioning with continuation between segments.

How it works

Per segment it builds the same joint audio+video latent the official nodes do, samples it in a single stage (KSampler plus MiniMaxH3SigmaShift), and decodes through LTXVSeparateAVLatent, which splits the joint latent into the video and audio halves. Then comes an optional second pass - 二采 - either a low-strength H3 resample or the LTX 2.5 two-stage refine (euler_ancestral, CFG 1, a fixed low-noise sigma tail, first-frame guidance, audio latent untouched). It has a quality gate: if the second pass doesn't measurably sharpen, you get the first-pass result instead of mush.

That's what the two comparison outputs are: compare_p1 is the first pass, compare_p2 is the refined/final one. With refinement off, they're the same frames.

Model loading is internal too: a card inside the node points at your UNET, reference UNET, CLIP, video VAE, audio VAE, Turbo LoRAs and the optional LTX 2.5 refine stack. Prefer canvas loaders? The optional model, clip, video_vae, audio_vae, model_ref and ltx_* inputs override the card when connected.

The inputs that matter

Beyond task_type and global_prompt: total_frames defaults to 124 - about five seconds at 24 fps - and snaps to H3's 17k+5 grid. width/height default to 864×480, frame_rate to 24 because that's what H3 was trained at, steps to 25, cfg to 1 (the official template's value; don't raise it), sampler to res_multistep and scheduler to simple. shift_video (12) and shift_audio (3) are the SigmaShift values - leave them unless you're chasing a specific look.

Then the performance toggles, which are where the real decisions live. clear_vram_between_segments unloads between segments - keep it on unless you have VRAM to spare. sag_enabled switches attention to Comfy Kitchen's INT8 path, which the author notes is close to SageAttention in speed and memory and actually works on H3; SageAttention's int8 route is disabled on purpose because it collapses on H3's long sequences. preview_tae gives you real-colour frames mid-sampling, but only if taeh3.safetensors is in models/vae_approx - otherwise it silently degrades to the cheap latent2rgb preview, which is why people think the preview looks bad.

timeline_data is the internal string the UI writes. Don't hand-edit it.

Outputs: images (the finished frames, as a list), audio, fps, plus compare_p1 and compare_p2. The node is also an output node - it saves the muxed video itself through the same path as ComfyUI's CreateVideo + SaveVideo, so you can delete those nodes from your graph. Wire compare_p2 (and compare_p1), images, audio and fps into the pack's MinimaxH3DYTScVideoCompare if you want the before/after view.

Install

Same pack as the rest of these nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/792877530-star/ComfyUI-MinimaxH3DYTsc

Restart ComfyUI. Manager users: search minimaxH3DYTsc6.0. You need a recent ComfyUI (the pack calls comfy_extras.nodes_minimax_h3 and checks its signatures), the H3 UNETs, the Qwen3-VL H3 text encoder loaded as type minimax, the video VAE and minimax_h3_audio_vae_fp32.safetensors. The requirements are opencv-python-headless, imageio-ffmpeg, scenedetect and comfy-kitchen>=0.2.34. The LTX 2.5 refine stack is optional - skip the models and the second pass just doesn't happen.

Where people get burned

• The model card looks filled in and nothing loads. Every default slot ships with a filename but as an open circuit - you have to tick it to actually load. The pack even detects the "everything disconnected" state, restores the slots by filename and warns you, because that state can't generate anything.
• /minimax/director/* returns 404. The node's HTTP routes register lazily; if PromptServer wasn't up yet at plugin load you get a warning in the console and a broken panel. Restart ComfyUI.
• A reference segment dies on the Audio VAE. Same rule as everywhere in this pack: reference conditioning needs the audio VAE loaded in the model card.
• This is a UI-heavy node, so the ComfyUI frontend can break it. Nodes 2.0 - the Vue rewrite of the canvas - has been rough on packs that draw their own panels, and the lesson from rgthree is that a single maintainer may not port. If the timeline panel misbehaves, the legacy canvas is still the safe fallback.
• GGUF checkpoints need ComfyUI-GGUF installed and a restart. And many of this pack's error strings are in Chinese: accurate, but not searchable.