ComfyUI-Hillobar

ComfyUI-Hillobar

Custom nodes for ComfyUI. Everything registers under a single top-level Hillobar menu category.

Nodes

MiniMax H3 Progressive Sampler (staged)

Progressive-resolution sampling for MiniMax H3. Early flow-matching steps only settle coarse structure, so they do not need full token density — this node runs the first part of the schedule on a smaller video latent grid, upscales the x0 estimate at each stage boundary, and finishes at the target resolution.

Video tokens scale with latent_t * (h // 2) * (w // 2), so a 0.5x spatial scale is roughly 4x fewer tokens: about 16x cheaper attention and 4x cheaper MLP for the steps that run small. End to end the win is bounded by how much of the schedule runs small — 55% of steps at 0.5x lands near 1.6–2x, not more.

Each stage is its own guider.sample() call, so conds, the packed layout, latent_shapes and the VRAM reservation are rebuilt per stage. No ComfyUI internals are mutated.

Inputs — guider, sampler, sigmas, latent_image, noise_seed, schedule, upscale_method, verbose. Outputs — output, denoised_output (both LATENT).

Feed it a latent already at the final resolution; it shrinks internally. That keeps the VRAM estimate an over-reservation rather than an under-reservation.

schedule

Comma-separated scale:end_percent stages, e.g. the default 0.5:0.55, 1.0:1.0 — run 55% of the steps at half the spatial latent grid, then finish at full size. scale is the spatial latent scale, end_percent is where that stage ends as a fraction of the step count. The schedule must end with 1.0:1.0. Stages may not shrink the grid mid-run. Use 1.0:1.0 alone as an A/B baseline.

Supported conditioning

t2va and ref2va. Reference blocks carry their own latent_h / latent_w, so they stay self-consistent when the target grid changes.

Not supported: fl2va keyframes (their cond rows share the target spatial grid while the cond latents stay full size) and noise masks (packed against the input latent shape, so they cannot follow a resolution change). Both raise a clear error rather than producing a silently wrong result.

Installation

Clone into ComfyUI/custom_nodes/ and restart ComfyUI. No dependencies beyond what ComfyUI already installs.

Structure

ComfyUI-Hillobar/
├── __init__.py                            # re-exports the node mappings
├── nodes/
│   ├── __init__.py                        # merges each module's mappings
│   └── minimax_h3_progressive_sampler.py
├── pyproject.toml                         # Comfy Registry metadata
├── requirements.txt
├── LICENSE
└── README.md