MiniMax H3 Progressive Target-Sparse Continuum [Experimental]

Sometimes "[Experimental]" in a node name is marketing. Here it's the author waving both hands. MiniMax H3 Progressive Target-Sparse Continuum is the rare node whose own README tells you not to use it for production output: real decoded-media testing showed cascading quality errors - skin imperfections, odd clothing changes, spurious background additions that propagate through later continuation chunks. It exists as a research and control path, not as the thing you render with. If you landed here googling "fast H3 continuation," you want its sibling in the same pack instead.

What it actually does

All the Continuum progressive nodes in this pack share one goal: make chunked exact-prefix continuation cheaper by not doing every early step at the full target grid. Mixed-Grid does that by generating the suffix on a real lower-resolution grid and upscaling it with a learned model. Target-Sparse tries a different bet: keep the sampler latent on the target grid the whole time, and instead sparsify only the early H3 hidden-token stream over generated video rows - retaining every protected video row plus a coarse target-grid anchor lattice, then restoring the full hidden grid before H3's native final layer.

That's an elegant idea and it's the point of the node for research: it isolates whether hidden-space sparsification alone can buy speed, without the added machinery (and quality) of a learned latent upscale. On paper it's attractive. On decoded media, the author reports, it fell apart - cascading artifacts that get worse as continuation proceeds, precisely because there's no learned upscale to fix what the coarse anchor lattice lost.

Inputs that matter

The required inputs mirror the Mixed-Grid node, which tells you a lot about intent - this was built for apples-to-apples comparisons:

• model and trajectory (an H3_FLOW_TRAJECTORY from the pack's MiniMax H3 Flow Trajectory node, one per run).
• source_mode / source_scale / source_width / source_height - and here's a subtle trap: on exact-prefix chunks these control anchor density, not sampler latent geometry. The sampler latent stays on the target grid; you're choosing how coarse the early hidden-token anchor lattice is, not rendering small.
• handoff_coordinate (0.35) and guidance_mode (direction is the conservative default; the other modes are research levers).
• handoff_transfer - bicubic (default, preserves the released handoff) or learned_3d (one connected H3_LATENT_UPSCALER provider applied at the exact-probe clean state).
• suffix_dc_bridge (default on) - the Continuum-only one-token per-channel DC seam correction. Here it calibrates from the first actual full-grid H3 predicted-clean boundary (no Spectrum forecasts), fixes only the first generated suffix token, and leaves the authoritative prefix untouched.

Outputs are the patched model and a metrics stream (H3_FLOW_METRICS) for the pack's diagnostic nodes - which is what you're supposed to be looking at if you're using this node.

Installing it

Same pack, same install as the others:

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/MiniMax-H3-Flow-Aligned-Regenerate.git

Restart ComfyUI (or install "MiniMax H3 Flow-Aligned Regenerate" from ComfyUI Manager). No extra Python deps, no model downloads from this pack - it patches the H3 sampling stack you already have. If you want the learned_3d transfer option, clone the companion MiniMax H3 Latent Upscaler pack too; with bicubic you don't need it. It's from xmarre, the same dev behind the well-regarded Spectrum and DiffAid research ports, so the engineering is careful - this is a deliberate negative result being shared, not a broken node.

Using it honestly

If your goal is output, don't. The failure mode isn't a subtle quality dip - it's a cascade: small artifacts in one chunk become bigger problems in the next, and because the pack doesn't do the learned upscale on this path, nothing rescues them. The author's release recommendation is Mixed-Grid Continuum for accelerated exact-prefix continuation, and this node's README position is unambiguous about why.

Where it is useful: architectural experiments, controlled comparisons, and validation baselines - e.g., measuring how much of the Mixed-Grid win comes from the learned transfer versus the low-grid sampling itself. If you run it, judge with decoded media, not structural metrics alone, and keep the suffix_dc_bridge on so the boundary seam at least isn't muddying your comparison. It's a rare honest negative result in a corner of the ecosystem where everyone else is selling you their experimental node as production-ready. That's the value here - just don't render your next video with it.