MiniMax H3 Chunked Two-Pass Plan (Advanced EXP/T8)

The classic two-pass recipe from image generation - render small, upscale, re-sample at low denoise - has a video-shaped problem: your "image" is a joint audio-video latent that gets very long, very fast. This node is the planning half of the T8 pack's chunked two-pass answer for MiniMax H3. It produces a plan describing how a learned 3D latent upscale and a temporal-chunk H3 re-sample should run. The other half, MiniMaxH3ChunkedTwoPassUpscaleT8Advanced, executes it.

The idea is straightforward and worth knowing before you touch any slider: first pass generates a low-resolution H3 clip, then an independent learned 3D upscaler model grows the latent (not the pixels) to high resolution, and then H3 re-samples at the higher canvas to add real detail. That's different from plain interpolation, and the KB's upscaling notes back the intuition: upscaling the intermediate and doing a low-denoise second pass is the standing recipe for coherent high-res output. The "learned" part is the difference between this and the pack's simple 32-divisible latent upscale - a model trained to add plausible detail instead of just resizing.

The inputs that matter

There are a lot of sliders here, but the safe defaults are actually safe. The ones you'll touch:

• model_name - the learned 3D latent upscaler from ComfyUI/models/latent_upscale_models (the pack expects minimax_h3_latent_upscaler_3d_fp16.safetensors).
• target_width / target_height - default 1280×704. The note in the description is worth reading: there is no project pixel-area ceiling; memory and runtime are yours to own. 1920×1088 is a risk reference area, not a hard limit, and bigger will warn but not block.
• temporal_chunk_frames (136) and temporal_overlap_frames (17) - the clip is split into temporal chunks so the whole thing fits in memory. The overlap lets chunks blend into each other rather than showing seams.
• anchor_strength (0.999) - how strongly each chunk is anchored to the previous one's result. This is your seam-control knob.
• spatial_strategy - the one that decides everything about risk. full_frame_safe keeps each temporal chunk full-frame, which preserves H3's global spatial context. independent_tiles_exp splits the frame into independent spatial canvases to save memory, and the "exp" suffix is doing real work: those tiles' content can diverge, so it's research-only.

Tiles, overlaps, fades, blend mode, precision and release policy round out the plan - leave them alone until you know why you're changing them. Outputs are the plan object (feed it to the Upscale executor) and a report_json describing what you're about to do.

Who it's for

If you're generating H3 at 1472×832 or bigger on a 16GB card and hitting OOM, this is the workflow's planning stage. It pairs naturally with the pack's 13-latent-upscale workflows, which use a low-res 4-step pass + learned upscale + high-res pass. And keep the pack's honesty rule in mind: chunked two-pass doesn't mean "unlimited resolution is now free." It lowers the peak memory by chunking the temporal dimension; it doesn't lift the per-chunk cost.

Install

Same pack, same steps:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

or ComfyUI Manager → "MiniMax H3 Audio T8" → restart. Update ComfyUI itself first. The one model you do need for this node is the learned 3D latent upscaler - grab it from the pack's model links or the t8star HuggingFace org and drop it in ComfyUI/models/latent_upscale_models.