GUIDED AV LOOP · EXPERIMENTAL · One Master AV Latent

Every other long-video approach in this pack, and in the ecosystem generally, has a moment where two clips have to be joined: decode, trim the overlap, blend, hope. The Guided AV Loop sampler is the one that doesn't. Its technical windows overlap inside a single master AV latent, so there is no decoded join, no trim and no crossfade between chunks. Continuity isn't repaired at the boundary because there is no boundary in the latent.

That's the whole pitch, and it's why the node is worth the friction.

Inputs

Required: model, positive, latent, video_vae, audio_vae, cine_linx, chunk_index, seed (42), seed_stride (1), steps (20), sampler_name and scheduler (populated from your ComfyUI's own sampler list, with res_multistep and simple moved to the front), denoise (1.0), shift_video (12.0) and shift_audio (3.0).

positive and latent come from IAMCCS_MiniMaxH3MaskedLoopConditioning. The separate video/audio sigma shifts are how H3 tunes the two modalities independently, and the defaults here are the ones the pack uses elsewhere.

motion_state is optional and explicitly ignored - kept so that graphs built for the standard backend keep validating when you switch modes.

Outputs

native_frames, native_audio, bridge_last_frame, sampled_latent, native_fps, report. Straight into a preview strip, a save, or the lazy router's guided_* sockets.

What it insists on

This node has the strictest precondition list in the pack, and each one is a real guard rather than bureaucracy:

• Task mode must be guided_av_loop_experimental. It won't quietly run under a different mode.
• chunk_index must be 0, and the plan must hold exactly one chunk. Guided AV Loop is one master-latent execution. If your plan has several chunks, you want Continuous AV instead.
• ComfyUI needs continuous per-row MiniMax H3 mask support (issue #15375). Without it the loop provider can't do the in-place windowing and the node refuses rather than producing something with a seam.
• No progressive-spatial acceleration. Masked windows and progressive sampling don't mix, so pick Native, PDD, FastH3, SLA or a compatible Turbo LoRA. Fused Fast preview is also excluded from this branch.
• Per-window prompts, or nothing. If the plan has multiple technical windows and it received one shared prompt, you get Use IAMCCS MiniMax H3 Masked Loop Conditioning before this sampler; a shared prompt would leak later framing directions into earlier windows. That error message is a small essay on why the conditioning node exists.

It also runs the pack's Turbo/FastH3/PDD/secondary LoRA stack and the audio-lock audit before sampling, so acceleration settings from H3 Settings still apply - just not the progressive ones.

Being honest about the cost

The trade for "no join at all" is that everything happens in one latent, and that latent is big: several overlapping windows of a full AV take, sampled together. This is the mode where a card with headroom earns its keep, and where lowering steps or picking a Turbo path actually matters rather than just being a preference.

It's also labelled EXPERIMENTAL in its own display name, which in this pack is not decoration. If you're on a deadline, Continuous AV or LatentGoAhead will give you a finished film with joins you can inspect and fix. If you're chasing a genuinely seamless long take and you're willing to iterate, this is the shortest path to it that I've seen in node form.

Install

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

Manager → IAMCCS → install → restart. On top of the usual H3 stack (FL2VA model, text encoder, video VAE, audio VAE, Shotboard plan on cine_linx), check your ComfyUI build for the per-row mask feature before you spend an evening authoring a programme this can't sample. The failure is clean, at least - you'll know within seconds.