MiniMax H3 Native Source Continuation

The pack this node lives in is a research release: a dozen experimental handoff variants, a wall of deprecated controls, and this one at the top of the README labelled "recommended." If you want the short version - MiniMax H3 Native Source Continuation is the progressive-generation node the author actually stands behind.

Think of it as hi-res fix for video. H3 is a 33B omni-modal model - video and native stereo audio generated together, up to 2K - and the expensive part is sampling the whole clip at target resolution. Progressive generation spends the early part of one schedule on a smaller grid, where motion and composition are cheap, then hands off to the target grid for detail. One schedule, one transition.

How it works

It's a MODEL-patching node. You feed it the loaded model plus a trajectory handle (trajectory, from MiniMax H3 Flow Trajectory), and it returns a patched model you wire into both the scheduler and the guider.

Underneath, the low/probe stage samples one full-duration clip on the reduced source grid. The learned 3D upscaler named by learned_upscaler transfers that trajectory to the target grid, and high refinement runs there using the original target prefix. Then the audio/video contract: protected prefixes stay exact, and the node adds no extra H3, provider or VAE calls to do it.

The interesting part is native source carry. For a continuation, the node reuses the preceding chunk's already-computed source clean prediction, paired against the exact returned target suffix - the pairing is a float32 hash match, not a similarity check. That's what makes the handoff cheap - nothing is recomputed to rebuild the source context. And there's no silent fallback: missing or mismatched context stops before low sampling rather than quietly substituting a reconstruction.

The inputs worth setting

You can leave most of this panel alone. The defaults are the shipped example: source_mode=scale, source_scale=0.70 (with source_width/source_height at 864/640 sitting idle until you pick pixels mode), handoff_coordinate=0.35 with handoff_selection=fixed, and guidance_mode=direction+temporal.

Two traps in that panel. acceleration_weight and consistency_weight are staged values - with guidance_mode=direction+temporal the runtime simply doesn't use them. They only become live in direction+acceleration and downsample_consistency respectively, so tweaking them and seeing nothing change is expected behaviour, not a bug. And source_prefix_projection defaults to native_source_carry; the alternative, latent_bicubic, reconstructs the source prefix from the target prefix and is the escape hatch when paired native context is unsuitable. Neither mode modifies protected target bytes.

Also worth knowing: learned_upscaler is a required input, not a dropdown - without a connected H3_LATENT_UPSCALER provider the learned path has nothing to call. audio_guided_overlap_mode defaults to sampler_mask_exact_timestep, which protects carried audio; in that mode audio_guided_overlap_ticks (16 by default) has zero effective release, so don't tune it and expect a result. uniform_source_detail_transport defaults on: it transfers a last-prefix exact-minus-learned detail residual before high refinement, adds no model or VAE call, but costs compute and memory during calibration and reports a skipped correction if its own checks don't pass.

Optional extras: metrics, and capture_boundary_witness, which saves stage snapshots for the offline audits and adds copy and disk-write time per run.

Outputs are model and metrics. The model goes to your scheduler and guider; metrics goes to MiniMax H3 Metrics JSON, which writes the run's structured model/sampler/geometry record to disk.

Install

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

ComfyUI Manager is the intended route - search MiniMax H3 Flow-Aligned Regenerate. The pack's own Python dependency list is just numpy; the weight is in the companions, and this node needs two of them. VDN-H3-Plus supplies the partitioned attention transport the uniform-source path requires (Sol-H3 is optional, not a prerequisite), and the learned upscaler comes from Comfyui_Minimax_h3_latent_Upscaler-Plus - its provider node with minimax_h3_latent_upscaler_3d_bf16.safetensors, device cuda, precision bf16, offload_after_upscale=false. For a sequence you also want H3 Continuum-Plus. Use a recent ComfyUI build; the pack's release notes point at a Core adapter repair PR for INT8 fused MLP runtime adapters.

Where this bites people

The number one mistake is running part of a sequence with a different model. Native carry is hashed against the preceding chunk's returned suffix, so a cached first chunk generated with another MODEL cannot supply it: keep the same patched MODEL for the initial chunk and every continuation.

Then the geometry contract: batch-one H3 video, native 5k+2 latent-token lengths, an even spatial grid, and a target larger than the reduced source grid. Chunk indices have to advance consecutively, and the prefix has to end on the same native phase. Finally - and this catches people after a successful run - any latent processing between generation and continuation that changes the carried target bytes will fail the pairing check, including post-sampling refinement applied to the prefix. Refine upstream, not in the middle.

Last thing, straight from the README: slight continuation darkening is a known issue, and the defaults don't guarantee seam quality or timing for every prompt and reference. Judge on decoded media.

And one bit of context if you googled H3 rather than this pack: the weights ship under the MiniMax H3 Community License, whose Applicable Territory excludes the US, EU, UK and South Korea - not a technical problem, but worth knowing.