中文说明

Progressive-resolution sampling for ComfyUI: run the early denoising steps at low resolution, lift the result to full resolution, and finish there — faster generation with no training. Based on the SelfLift paper (SelfLift-zero) for rectified-flow image models, plus an experimental MiniMax H3 audio-video adaptation that the paper does not validate. Also included: an experimental H3 port of TST temporal-attention correction.

All nodes are in the selflift category. The samplers use sampler/sigmas inputs like SamplerCustom; connect the standard euler sampler from KSamplerSelect and the model's normal scheduler (other samplers are rejected).

The paper's SelfLift-zero for rectified-flow image backbones. Connect 4D image latents (e.g. Empty Latent Image) and the model's own VAE.

The transition adds no denoiser evaluations: an N-step schedule stays exactly N NFEs, plus one VAE decode → upscale → re-encode round trip unless rho=0.

Experimental H3 audio-video adaptation (not paper-validated). Two modes:

• External upscaler (default): an installed H3 checkpoint with rho=0 — a learned latent-only lift. Practical default, but not SelfLift-zero.
• SelfLift-zero: upscaler_model=none with rho>0. Suggested H3 starting point: rho=0.6, w_min=w_max=1 (see Diagnosing H3).

Extra inputs: upscaler_model, upscaler_unload (default on: unloads the upscaler from VRAM right after the lift, before the high-resolution stage) and highres_tiling (experimental: splits the high-res phase into 1–8 spatial tiles to save VRAM; only the first tile's audio is kept, there is no cross-tile attention, ControlNet is unsupported, and quality/speed may change).

Download from LBH-123-AI/Minimax_h3_latent_Upscaler into ComfyUI/models/latent_upscale_models/ and restart. The node selects the first filename containing h3, otherwise none (nearest-neighbor lift).

A MODEL → MODEL patch node that improves the temporal stability of H3 videos at inference time — no training, no extra model, ~1–2% runtime overhead. Typical problems it counteracts: details that flicker or morph between frames (logos, on-screen text, textures), identity drift of people and outfits, and physically implausible motion. It works with any standard sampler node, not only the SelfLift sampler. It cannot create detail the model does not have.

How it works, in one paragraph: instead of blindly strengthening cross-frame attention, TST measures a signed "Spectral Tension" of the frame-level attention operator and corrects only imbalanced heads — sharpening over-mixed ones, softening fragmented ones — with stronger correction in deeper layers and earlier steps. Theory and derivations are in the paper.

The diagnostics line reports latent frames, grid rows per frame, tension before/after the within-call correction, mean gamma, the share of corrected heads, and TST's own runtime. On real H3 runs the correction direction matched the exact attention operator for 89–100% of heads. Set SELFLIFT_TST_EXACT=1 before starting ComfyUI to log the exact-operator comparison on one layer per forward (debug only, slow).

Limitations: MiniMax H3 models only (other models are skipped silently); not compatible with highres_tiling (skipped with a console warning).

Same prompt and seed, transition_step=6, lowres_scale=0.5:

A controlled single-seed H3 run found: clean native baseline and clean pure pixel anchor; widespread artifacts on the nearest route that the paper's image setting (rho=0.3) mostly left visible; rho=0.6 with w_min=w_max=1 removed the main artifacts. Single-seed evidence only — H3's direct-lift error is broader than on the paper's image backbones, so start strong and reduce only if the result is overly smooth.

• [SelfLift plan] — resolved low/target latent shapes, NFE counts, transition sigmas, enabled lift routes
• [SelfLift timing] — wall-clock time per step and per stage (low-res, transition, high-res)
• [SelfLift upscaler] / [SelfLift tiling memory] — upscaler inference and tiling memory estimates (heuristics, not measured peaks)
• [H3 TST] — TST per-forward diagnostics
• SELFLIFT_TIMING_SYNC=1 — CUDA-synchronized timing (slower; for diagnostics)
• SELFLIFT_MEMORY_LOG=1 — host/device memory snapshots at stage boundaries
• SELFLIFT_DEBUG=1 — dump transition intermediates as PNGs under debug/ (slow, memory-hungry)
• SELFLIFT_TST_EXACT=1 — TST exact-operator calibration probe (debug only)

• Only standard Euler with s_churn=0.
• noise_mask is supported on both samplers with Set Latent Noise Mask semantics (1 = generate, 0 = keep the original content; use an initialized latent to have content worth keeping). Masks at any resolution are resized to the latent grid, and a time length of 1 is shared over all frames. The keep-region is pinned to the original latent at every step, and the artifact-aware correction is restricted to the generate region. noise_mask is not compatible with highres_tiling.
• Use the VAE belonging to the sampled model so the pixel anchor stays in the same latent space.
• H3's 768-pixel short edge becomes 384 px at lowres_scale=0.5, which may sit outside the backbone's training distribution — validate per model.
• SelfLift-rich (the distilled lifter + On-Policy Self Recovery) requires training and is not included.

• SelfLift paper: SelfLift: Accelerating Few-Step Diffusion via Self-Recovering Resolution Transition
• TST paper and code: Temporal State Transport in Video Generation, lytang63/temporal-state-transport
• Optional MiniMax H3 latent upscaler checkpoint and download: LBH-123-AI/Minimax_h3_latent_Upscaler
• Original ComfyUI integration and inference implementation: LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler

Thanks to LBH-123-AI for publishing the MiniMax H3 latent upscaler weights and ComfyUI implementation. They made the optional learned H3 lifting path in this plugin possible. This external lifter remains separate from the SelfLift paper's SelfLift-rich model.

@article{wen2026selflift,
  title={SelfLift: Accelerating Few-Step Diffusion via Self-Recovering Resolution Transition},
  author={Wen, Tingyan et al.},
  journal={arXiv:2609.02036},
  year={2026}
}