You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

darksidewalker/MiniMaxH3

Official repository for the GRAVITON weight-space fusion architecture and merged MiniMax H3 audio-video DiT checkpoints.

  • GRAVITON Architecture, Mathematical Specification & Engine Author: Darksidewalker
  • Official Reference Implementation: darksidewalker/dasiwa-quant-station (core/model_merge_engine.py, commit 69412c0+)
  • Engine & Workbench License: GPL-3.0-or-later
  • Specification & Whitepaper License: CC-BY-SA-4.0 (Attribution + ShareAlike required)

Evolution of MiniMax H3 Merge Architectures

This repository encompasses three generations of model-level H3 fusion recipes implemented in DaSiWa Quant Station. While earlier approaches targeted specific conditioning unifications (fl2va + ref2va), GRAVITONβ€”designed and engineered by Darksidewalkerβ€”provides a mathematically rigorous, calibration-free spectral and polar geodesic framework for arbitrary DiT pairings.

Recipe Engine ID Primary Use Case Core Mechanism Key Limitation Solved by Next Generation
1. Hybrid H3 h3_hybrid Grafting ref2va conditioning onto fl2va Hard block/row splicing of AdaLN weights & biases (default blocks 30–49) + optional final_layer.audio_out swap. Zero interpolation; hard boundaries between block regimes; requires unquantized matching layouts.
2. Delta-Fused H3 h3_delta Single checkpoint serving both keyframe & reference paths Linear delta injection (W = fl2va + s Β· Ξ”) with optional fixed-rank (N) randomized SVD truncation on 2-D trunk matrices. Fixed rank ignores aspect ratios; linear interpolation on heterogeneous finetunes/turbos causes variance collapse.
3. GRAVITON h3_graviton Merging any two H3 models (Base, Turbo, Finetunes, W6A8) Full-spectrum RMT filtering (Gavish-Donoho) + pre-geodesic TIES/EMR sign mediation + row-wise polar SLERP & RegMean norm re-anchoring. Current architecture β€” prevents variance drift, spectral tail noise, post-SLERP trajectory warping, and AdaLN black-frame collapse.

1. GRAVITON (h3_graviton)

GRAVITON Event Horizon Pipeline

Geodesic Rank-Adaptive Variance-Isolated Tensor Orthogonal Normalization
Designed and authored by Darksidewalker.

Authorship & Architectural Lineage

While originally motivated by an unimplemented, four-step community recipe description (OliviaRossi/Minimax-h3_Singularity-x-10Eros-Max_PRISM, which had no published code or formal paper), GRAVITON is an original mathematical formulation and systems implementation by Darksidewalker (core/model_merge_engine.py, commit 69412c0+) that corrects fundamental geometric flaws in prior descriptions and introduces:

  1. Pre-Geodesic Sign Mediation Ordering: Moves TIES/EMR sign-conflict resolution before spherical interpolation so post-rotation sign flips never knock rows off the geodesic hypersphere.
  2. Exact Full-Spectrum RMT & Gavish-Donoho Auto-Rank: Replaces arbitrary fixed-rank cutoffs with aspect-ratio-adaptive optimal hard thresholding via exact Gram-matrix eigendecomposition.
  3. Analytical Marchenko-Pastur Quantization Floor (tau_quant) & Comfy-Kitchen W6A8 Bridge: Derives exact uniform step-noise variance from QuantStation W6A8 scale companions and blocks quantization grid staircases from entering the structural subspace ($L$) or diffuse residual ($R$).
  4. AdaLN Anti-Alignment Protection & Structural Preservation: Prevents the black-video variance collapse caused by per-row SLERP on anti-aligned AdaLN modulation weights or mismatched community coordinate tables.
  5. Decoupled Float Output: Emits clean floating-point artifacts (BF16/F16/F32) rather than risking inline quantization packer corruption.

Why GRAVITON?

Merging step-distilled accelerators (8-step Turbo/DMD/LCM), diverse community finetunes, or group-quantized checkpoints into base DiT models triggers four primary failure modes:

  1. Spectral Tail Contamination: Step distillation concentrates structural trajectory changes into top singular vectors while leaving discretization artifacts in the spectral tail. Fixed-rank truncation (e.g., rank-64) under-fits wide layers and leaks noise from narrow projections.
  2. Post-Geodesic Trajectory Warping: Applying element-wise sign mediation after spherical interpolation flips components off the hypersphere and destroys the interpolated angle.
  3. Timestep Conditioning Instability: Scaling deltas by Ξ± > 1.0 or applying per-row SLERP inside adaptive layer norm (adaln) weights causes severe variance drift across denoising steps, resulting in saturated or black-video outputs.
  4. Quantized Manifold Mismatch: Group-quantized checkpoints carry Hadamard-rotated (ConvRot), bit-packed coordinates whose discretization noise inflates the delta spectrum.

Mathematical Pipeline

Given base weights $A \in \mathbb{R}^{m \times n}$ and target weights $B \in \mathbb{R}^{m \times n}$ with merge strength $t \in (0, 1]$ (extrapolation $t > 1$ is strictly refused):

Stage 0 β€” Comfy-Kitchen W6A8 Ingestion & Dequantization Bridge (Opt-In)

Fail-closed by default (accepting only F16/BF16/F32). When allow_quantized_inputs is enabled, the engine recognizes QuantStation's exact W6A8 header contract (I8 packed .weight, F8E4M3 .weight_s_rel, F32 .weight_s_channel, and per-layer .comfy_quant JSON) and dequantizes via comfy_kitchen's AsymW4A8Int8Layout.dequantizeβ€”cleanly inverting the ConvRot block-Hadamard rotation ($Q = Q^T = H_b / \sqrt{b}$). From the scale tensors $S = s_{\text{channel}} \cdot s_{\text{rel}}$, the analytical per-element uniform step noise variance is recorded: Οƒq2=mean(S2)12\sigma_q^2 = \frac{\text{mean}(S^2)}{12}

Stage 1 β€” Full-Spectrum RMT Split with Analytical Quantization Floor

For SVD-eligible 2-D trunk matrices, the weight delta $\Delta = B - A$ is decomposed via the exact spectrum of the smaller Gram matrix ($G = \Delta^T \Delta = V \Lambda V^T$, $\sigma_i = \sqrt{\max(\lambda_i, 0)}$): Ο‰(Ξ²)=0.56Ξ²3βˆ’0.95Ξ²2+1.82Ξ²+1.43,Ξ²=min⁑(m,n)max⁑(m,n)\omega(\beta) = 0.56\beta^3 - 0.95\beta^2 + 1.82\beta + 1.43, \quad \beta = \frac{\min(m,n)}{\max(m,n)} Ο„GD=Ο‰(Ξ²)β‹…Οƒmed(Gavish-Donoho optimal hard threshold)\tau_{\text{GD}} = \omega(\beta) \cdot \sigma_{\text{med}} \quad \text{(Gavish-Donoho optimal hard threshold)} Ο„quant=(m+n)ΟƒqA2+ΟƒqB2(Marchenko-Pastur bulk edge)\tau_{\text{quant}} = (\sqrt{m} + \sqrt{n})\sqrt{\sigma_{qA}^2 + \sigma_{qB}^2} \quad \text{(Marchenko-Pastur bulk edge)} Ο„=max⁑(Ο„GD,Ο„quant)\tau = \max(\tau_{\text{GD}}, \tau_{\text{quant}})

Singular vectors above $\tau$ form the low-rank structural update $L = (\Delta V_{\text{keep}})V_{\text{keep}}^T$, leaving the diffuse residual $R = \Delta - L$. The denoised target $B_{\text{raw}}$ is reconstructed with an energy-derived halo damping factor: Braw=A+L+Ξ³haloR,Ξ³halo=1βˆ’0.08β‹…capture(L)∈[0.92,1.0]B_{\text{raw}} = A + L + \gamma_{\text{halo}} R, \quad \gamma_{\text{halo}} = 1 - 0.08 \cdot \text{capture}(L) \in [0.92, 1.0] Whenever dequantization noise is present, $\gamma_{\text{halo}}$ is strictly capped at 0.15 so quantization grid staircases in $R$ never survive into the merge. (For float pairs, fixed-rank $N$ via Halko et al. randomized SVD remains available as an opt-out).

Stage 2 β€” Pre-Geodesic Sign & Interference Mediation

Executed on $B_{\text{raw}}$ before rotation: salient elements ($\vert{}B_{\text{raw}} - A\vert{} \ge \text{row median}$) whose sign conflicts with $A$ resolve to the magnitude-weighted winner ($t \cdot \vert{}B_{\text{raw}}\vert{}$ vs. $(1-t) \cdot \vert{}A\vert{}$, following TIES/EMR logic) to yield $B'$. Performing mediation prior to SLERP guarantees the spherical trajectory remains unwarped.

Stage 3 β€” Polar Geodesic Interpolation (Row-Wise)

Directional rotation (short-path SLERP) is decoupled from log-space geometric magnitude interpolation per output row $i$: A^i=Aiβˆ₯Aiβˆ₯,B^iβ€²=Biβ€²βˆ₯Biβ€²βˆ₯,ΞΈi=arccos⁑(clamp(⟨A^i,B^iβ€²βŸ©,βˆ’1+Ο΅,1βˆ’Ο΅))\hat{A}_i = \frac{A_i}{\Vert{}A_i\Vert{}}, \quad \hat{B}'_i = \frac{B'_i}{\Vert{}B'_i\Vert{}}, \quad \theta_i = \arccos(\text{clamp}(\langle \hat{A}_i, \hat{B}'_i \rangle, -1+\epsilon, 1-\epsilon)) ui=sin⁑((1βˆ’t)ΞΈi)A^i+sin⁑(tΞΈi)B^iβ€²sin⁑(ΞΈi)u_i = \frac{\sin((1-t)\theta_i)\hat{A}_i + \sin(t\theta_i)\hat{B}'_i}{\sin(\theta_i)} Wi=uiβˆ₯uiβˆ₯exp⁑((1βˆ’t)ln⁑βˆ₯Aiβˆ₯+tln⁑βˆ₯Biβ€²βˆ₯)W_i = \frac{u_i}{\Vert{}u_i\Vert{}} \exp\big((1-t)\ln\Vert{}A_i\Vert{} + t\ln\Vert{}B'_i\Vert{}\big) (Falls back to linear interpolation when $\theta_i < 10^{-4}$; negates $\hat{A}_i$ for the short path when $\cos\theta_i < 0$).

Stage 4 β€” Magnitude Re-Anchor & Decoupled Float Export

Each row $W_i$ is rescaled to the geometric norm target of the original, pre-denoised parent pair $\Vert{}A_i\Vert{}^{1-t} \cdot \Vert{}B_i\Vert{}^t$ (RegMean-style norm preservation). Merged weights are always emitted as clean floating-point tensors (BF16/F16/F32) and handed to downstream conversion jobs if packed W6A8 output is desired, eliminating inline packer corruption.


Tensor Family & Structural Preservation Policy

GRAVITON routes parameters strictly by architectural role:

  • 2-D Trunk Matrices (attn, mlp, out, proj): Full 4-stage GRAVITON pipeline (Spectral Denoise β†’ Sign Mediation β†’ Polar SLERP β†’ Norm Re-Anchor).
  • AdaLN Modulation Weights (ada): Spectral Denoise + purely linear delta (W = A + t Β· Ξ”_denoised). Because fl2va and ref2va AdaLN rows are strongly anti-aligned, per-row SLERP induces black-video collapse.
  • Exact 1-D & Conditioning Families (adaln_t_table / time_embedder, norms, biases, rope.inv_freq): Exact linear interpolation W = A + t(B - A).
  • Structural-Layer Preserve (preserve_structural=True by default): Protects conditioning-critical keys (adaln_t_table, time_embedder.*, adaln_proj.linear.*, per-block/qk norms, rope.inv_freq, final_layer.*, token_refiner.*, and patch/conditioning projections) by copying them directly from Base (A) when merging checkpoints with custom coordinate-system folds.

Recommended Operational Regimes (h3_graviton)

Pairing Anchor (Base A) Target (Overlay B) Recommended $t$ Expected Behavior
Base + Turbo (Balanced) Base bf16 8-step Turbo 0.45 – 0.55 Preserves base generative diversity with faster step convergence.
Turbo + Base (8-Step Lock) 8-step Turbo Base bf16 0.20 – 0.30 Locks the 8-step distillation schedule while injecting base textures/detail.
Base + Base (Finetunes) Finetune 1 Finetune 2 0.30 – 0.70 RMT thresholding strips uncorrelated overfitting noise between finetunes.
Turbo + Turbo Turbo 1 Turbo 2 0.40 – 0.60 Combines distilled aesthetics with zero AdaLN variance drift ($t \le 1.0$).
Float + W6A8 Base bf16 QuantStation w6a8_g32 0.35 – 0.50 MP quantization floor blocks grid artifacts from entering $L$; outputs clean bf16.

2. Earlier Approaches (h3_delta & h3_hybrid)

Delta-Fused MiniMax H3 (h3_delta)

H3 DELTA FUSED

Fuses the full ref2va βˆ’ fl2va weight delta back into an fl2va base so a single checkpoint serves both keyframe (fl2va) and reference (ref2va) conditioning across both pruned (adaln_t_table) and full (time_embedder) variants: W=Wfl2va+sβ‹…(Wref2vaβˆ’Wfl2va)W = W_{\text{fl2va}} + s \cdot (W_{\text{ref2va}} - W_{\text{fl2va}})

  • Rank 0 (Exact): Applies the full uncompressed delta across all tensors (~38 GB output).
  • Rank N (SVD-Capped): Applies randomized SVD capped at rank $N$ (e.g., $N = 1024$) to 2-D trunk matrices while keeping norms, biases, timestep tables, and RoPE frequencies exact.

Hybrid MiniMax H3 (h3_hybrid)

HYBRID H3

A zero-interpolation block/modality surgery recipe that grafts ref2va AdaLN modulation onto an fl2va base:

  • Whole-Block Mode: Copies entire ref2va AdaLN weights and biases into a selected block range (default blocks 30–49 of the 50-block grid).
  • Modality-Row Mode (Experimental): Independently splices ref2va video, audio, and text AdaLN rows across user-selected block grids, with optional final_layer.audio_out weight/bias replacement.

3. Regenerate 2K β€” Open-H3 refiner with native ComfyUI keys

Detail crop: draft upscaled 2x (left) vs Regenerate 2K native detail (right)

NVlabs' Open-H3-Regenerate-2K refiner LoRA converted to native MiniMax H3 key names so ComfyUI's LoRA loader actually applies it. The upstream file uses Diffusers key names, which ComfyUI silently skips key by key β€” the original file loads and does nothing. These files load with zero unloaded keys.

File Contents
lora/open_h3_regenerate_2k_native_bf16_comfyui.safetensors Exact key-converted merge (Open-H3 refiner rank-384 + LightX2V FL2VA Turbo v1.2 + LMS Sharp v1.0), 8.77 GiB
lora/open_h3_regenerate_2k_native_svd_e0999_comfyui.safetensors Same merge, SVD re-ranked at 99.9 % energy β€” 2.35 GiB; same-seed output measured 39.7 dB PSNR from the exact file, detail looks the same. This is the one to use.
lora/open_h3_regenerate_2k_native_svd_e0999_recipe.txt Full conversion + SVD re-rank recipe (energy 0.999, median relative error 3.15 %) that reproduces the file above

It takes a 768P H3 draft and regenerates it at 2560x1440 in three sampling steps (Euler, CFG 1, sigmas 0.909375, 0.857143, 0.666667, 0, LoRA strength 1). Turbo and Sharp are already merged in β€” load this on the plain H3 model and do not stack your Turbo LoRA again. The draft's audio is returned unchanged.

Requirements: DaSiWa custom nodes >= 0.5.49, node MiniMaxH3 β€’ Regenerate 2K (ComfyUI-DaSiWa-Nodes), and the H3 latent upscaler minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors in models/latent_upscale_models. Supported out of the box in DEAS β€” Dark Engine AI Studio (darksidewalker/DEAS), no manual wiring needed there. The node pins the recipe above, handles 4K via an intermediate 2K upscale plus tiling, and patches only a model clone.

Tested on RTX 5090 (32 GB), FL2VA pruned bf16 base, 124 frames, 1344x768 -> 2560x1440: ~19 min untiled (slightly sharper, no OOM), ~26 min with spatial tiling; peak system RAM 21-25 GB. Three steps from high noise amplify tiny numeric differences, so same-seed results are not stable across settings β€” expected, NVlabs says the same. Start frames (I2VA) and Continuity joins stay coherent through the pass.

Examples (draft left, regenerate right)

Regenerate 2K vs the usual latent-upscale-plus-refine path:

Credits: refiner and merged LoRA by NVlabs / Efficient-Large-Model (Open-H3-Regenerate-2K). Files here are key-converted derivatives, no training or tuning; the LoRA remains subject to the MiniMax H3 Community License β€” check its terms, especially for commercial use.


License, Copyright & Attribution

Copyright Β© 2026 Darksidewalker. All rights reserved on original methodology, whitepaper, and merge engine architecture.

This repository operates under a split licensing model to protect original open-source mathematical and software contributions while respecting upstream foundation model weights:

  1. GRAVITON Engine & DaSiWa Quant Station Implementation (core/model_merge_engine.py):
    Licensed under the GNU General Public License v3.0 or later (GPL-3.0-or-later). Any software derivative, port, or integration of the merge engine must remain open-source under GPL-3.0-or-later and preserve full copyright and NOTICE attribution to Darksidewalker.
  2. GRAVITON Whitepaper, Mathematical Formulation & Documentation:
    Licensed under the Creative Commons Attribution-ShareAlike 4.0 International License (CC-BY-SA-4.0). Any adaptations or derivative specifications must credit Darksidewalker and be distributed under the same license.
  3. Merged Model Checkpoints (.safetensors):
    Merged weights are derivative artifacts of the MiniMax H3 base models and remain subject to the upstream MiniMax H3 Community License Agreement (minimax-h3-community-license-agreement, as applied to MiniMaxAI/MiniMax-H3) and its acceptable use policies. When redistributing checkpoints merged with GRAVITON, preserve the .safetensors header provenance keys (minimax_h3_graviton=baked and h3_graviton_* metadata).

Citation

If you use the GRAVITON merge architecture, the analytical W6A8 Marchenko-Pastur quantization floor, or checkpoints from this repository in your research, merge toolkits, or derivative models, please cite:

@software{darksidewalker2026graviton,
  author       = {Darksidewalker},
  title        = {{GRAVITON: Geodesic Rank-Adaptive Variance-Isolated Tensor Orthogonal Normalization (DaSiWa Quant Station)}},
  year         = {2026},
  month        = nov,
  publisher    = {GitHub / Hugging Face},
  license      = {GPL-3.0-or-later},
  url          = {https://github.com/darksidewalker/dasiwa-quant-station},
  howpublished = {\url{https://huggingface.co/darksidewalker/MiniMaxH3}},
  note         = {Reference implementation in dasiwa-quant-station (core/model_merge_engine.py, commit 69412c0)}
}

References & Grounded Analogs

  • Optimal Singular Value Hard Thresholding: Gavish & Donoho, The Optimal Hard Threshold for Singular Values is $4/\sqrt{3}$ (arXiv:1305.5870)
  • Randomized Low-Rank Matrix Approximation: Halko, Martinsson, & Tropp (arXiv:0909.4061)
  • DARE (Drop And REscale): Yu et al. (arXiv:2311.03099)
  • TIES-Merging & EMR-Merging (Sign Conflict Mediation): Yadav et al. (arXiv:2306.01708) / Huang et al. (arXiv:2405.17461)
  • Spherical Linear Interpolation (MergeKit): Goddard et al. (arXiv:2403.13257)
  • RegMean (Norm Preservation): Jin et al. (arXiv:2212.09849)
  • Upstream Community Inspirations: OliviaRossi/Minimax-h3_Singularity-x-10Eros-Max_PRISM (conceptual inspiration) and diffusers-modular/MiniMax-H3-Pruned-Ref-Delta-Fused-r1024
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Papers for darksidewalker/MiniMaxH3