IAMCCS_LongVid-v1 · Positioned Guides

MiniMax H3 makes 4-to-15-second clips. That's the model. Everything past fifteen seconds is engineering, and this node is where the engineering gets configured.

It's not a generator. It has no prompt field, no media inputs, and it doesn't sample anything. It's a settings surface: you tune it, it publishes one cine_linx output, and the Shotboard-connected H3 branches downstream read that bundle instead of each carrying their own private copy of your choices. The settings node is the technical control room; the Shotboard is the source of truth for the sequence; cine_linx is the wire between them.

This particular one is the narrow surface. The full H3 settings node has dozens more fields; LongVid-v1 publishes only the ones the long-video positioned-guides branch consumes, and hard-pins the rest off - no upscale pass, no RIFE, no face detailer, no secondary LoRA, no ControlNet. Fewer ways to be confused about which panel is lying to you.

The groups that actually matter

Output size and seed. width/height default to 640×384 - low on purpose, because this branch generates small and finishes big. seed plus seed_policy decides repeatability: fixed_per_generation holds one seed for the whole generation, fixed_per_chunk walks it per chunk, and seed_stride sets the step between them.

The sampling core. steps (16), sampler_name (44 options, euler default), scheduler (simple), denoise, and - H3-specific - shift_video (12) and shift_audio (3) as separate values. That's because H3 generates picture and stereo audio as one thing; audio shift is a real dial here, not an afterthought. acceleration is the other big one, 20 choices spanning native, Sage, SLA, Sol, spectrum and low-VRAM profiles. Start at native and only move when you have a reason.

Turbo and PDD. turbo_mode (off, early_8_10, ckpt500_6_8) with turbo_lora_name and turbo_strength; then pdd_lora_name and pdd_strength. Read those two tooltips before you touch them: they want a native Kijai *_Acc-8Step*_comfy LoRA matched to your loaded trunk (FL2VA vs Ref2VA), and the author's instruction on strength is blunt - keep it at 1.0, because a shape-changing PDD head bank needs full trained strength and the node fails clearly rather than quietly substituting a value.

Memory. text_encoder_device (gpu_auto, cpu_direct, auto, cpu_safe_12gb), h3_clipproj_profile (4B v3.1 default, 8B, or off), h3_clipproj_load_mode (dynamic vs streaming), performance_profile (14 presets by card class), and the h3_exact_* group - h3_exact_profile is an editable hardware baseline (xx60 124f/5.17s, xx70 209f/8.71s, xx80 294f/12.25s, xx90 362f/15.08s), h3_exact_chunk_rows (2048 is the conservative 8–12 GB value), h3_exact_precision_mode (default Preserve native, so INT8/W4A8 checkpoints stay in their released precision), plus QKV streaming and attention memory. That last one is worth quoting to yourself: lowering it does extra exact work - a safety fallback, not an approximate attention cache. vram_clean_before_decode frees memory before the VAE runs.

The long-video part. motion_context_window_frames defaults to 362 and the tooltip says why: 362 restores the proven contract, and smaller values are explicit memory trade-offs - you are buying VRAM with continuity. flf_continuity_mode, flf_continuity_tail_frames (22/39/56) and flf_continuity_audio control how first/last-frame continuity is carried. longvid_guide_window_policy picks the windowing strategy: standard_positioned is the ordinary path with positioned guides and no hidden latent-prefix carry, adaptive_latent_tail enables the narrow high-resolution windows with a latent tail, and legacy_auto exists to keep older saves behaving the way they used to.

Then the deep end: longvid_terminal_endpoint_mode offers ten terminal strategies, from hard_image (the legacy exact final-guide behaviour) through various pianosequenza_* latent-closure modes to pianosequenza_hd, a self-contained two-stage 0.5×→1.0× progressive solve. longvid_pianosequenza_2stage_enabled is authoritative one-stage generation when off - off means off, and leftover saved values can't override it. longvid_pianosequenza_stage_profile picks the two- or three-stage shape. And if you enable multi-stage, upscale_link_to_native + upscale_link_factor (default 640×384 → 1280×768) make the low→high pair part of the sample itself rather than a post-pass you bolt on after.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Manager → IAMCCS Nodes does the same thing. Restart, hard-refresh, one copy only. Then expectations: this node does nothing alone. It needs a working H3 setup (model, VAE, text encoder, audio VAE) and the Shotboard branch it feeds, and some acceleration options depend on Sage Attention or the Sol-Attn node being installed - picking one that isn't surfaces as a load-time failure, not a fallback. The pack ships no weights, and H3's own licence excludes the US, EU, UK and South Korea, outputs included.

How to actually use it

Change one group at a time. Set performance_profile and h3_exact_profile for your card, get a short single segment rendering, and only then start moving motion_context_window_frames, flf_continuity_* and the terminal mode. The defaults aren't arbitrary - they're the author's tested contract, and the long-video failures you read about in every community thread (drift, seams, clips that decelerate into their final frame) live in exactly those fields.