t8star/minimax_h3_turbo_4step_10ErosMax_test4_pruned_curveproj1025_T8

MiniMax H3 Turbo 4-Step — 10Eros Pruned Curve Compatibility LoRA (EXP v001)

An exact-checkpoint-specific, four-step ComfyUI acceleration LoRA for the local 10Eros_Max_h3_fl2va_bf16_test4_pruned.safetensors MiniMax H3 fine-tune.

This repository contains an independent LoRA, not a merged 40GB diffusion checkpoint. It keeps the 208 already compatible attention/MLP adapters from a legacy Standard MiniMax H3 Turbo LoRA and converts its 51 full-width AdaLN adapters to the target checkpoint's 8-dimensional pruned curve basis.

This is an Experimental, exact-base compatibility build. It is not an official MiniMax release, not a newly trained or re-distilled Turbo model, and not a generally compatible LoRA for arbitrary pruned checkpoints. Use it only with the exact main-model SHA-256 documented below.

MiniMax H3 and Model Derivatives are governed by the MiniMax H3 Community License Agreement. That agreement defines excluded territories, redistribution conditions, commercial terms, downstream safeguards, an acceptable-use policy, and mandatory LICENSE/NOTICE handling. A public Hugging Face repository can be reachable from excluded territories; repository gating is not necessarily geographic access control. Confirm that you have the right to distribute both this derivative and the 10Eros fine-tuned-base-dependent adaptation. This model card is not legal advice.

中文说明

这是为以下唯一指定主模型制作的 MiniMax H3 四步加速 LoRA：

10Eros_Max_h3_fl2va_bf16_test4_pruned.safetensors
SHA-256: f82cc3f723b080e7ae94a7c98f95aa989e387618d0bdc940133dfbd9f432c062

普通完整 MiniMax H3 Turbo LoRA 有 259 组 adapter，其中 208 组 attention/MLP 与该 pruned 主模型形状兼容，另外 51 组 AdaLN 的输入维度为 2688，而主模型只接受 8 维曲线坐标，直接用 Bypass 加载会发生维度错误。

本模型没有简单删除这 51 组 AdaLN，而是：

• 保留 208 组原兼容 adapter，416 个 tensor 与来源逐张量完全一致；
• 将 51 个 AdaLN-A 从 [16,2688] 投影为 [16,8] BF16；
• 51 个 AdaLN-B 保持逐张量完全一致；
• 增加 51 个 FP32 .diff_b，保留仿射投影中的常数响应；
• 最终为 259 组 A/B adapter + 51 个 bias diff，共 569 tensors。

Model details

Property	Value
Model type	MiniMax H3 ComfyUI acceleration LoRA
Target task	FL2VA-family joint video + synchronized stereo audio generation
Training	None for this conversion
Distillation source	Existing legacy Standard four-step Turbo LoRA
Target architecture	Exact 10Eros curve-pruned BF16 checkpoint
Recommended steps	4, the only locally validated setting
Precision	BF16 A/B adapters; FP32 .diff_b
Adapter count	259 A/B adapters + 51 bias patches
Tensor count	569
File size	794,888,696 bytes, approximately 758.06 MiB
Status	Experimental exact-checkpoint compatibility candidate

Weight provenance

The acceleration source is the exact local file:

File: minimax_h3_turbo_4步加速_comfyui.safetensors
SHA-256: 35946f9f2957c2766e28b627c88169535249dd07a3040ce3c2c8c99951fdbc7b

Its embedded conversion metadata points to local source SHA-256 9344cd958f8d354da03dd00b7d462933eb5d0cbf11e56a25d8e9911bb971160e. Independent provenance inspection identifies it as a small BF16/T8 perturbation of the initial/legacy non-EMA Standard four-step line from larryvrh/MiniMax-H3-Turbo-Lora.

It is not v4-step600, ckpt500, ckpt850, or EMA. Do not rename or advertise this file as one of those checkpoints. The current upstream repository may recommend newer weights; that recommendation does not change this file's provenance.

Exact compatible main model

File: 10Eros_Max_h3_fl2va_bf16_test4_pruned.safetensors
Size: 40,225,724,112 bytes
SHA-256: f82cc3f723b080e7ae94a7c98f95aa989e387618d0bdc940133dfbd9f432c062
Structure: 532 tensors; adaln_t_table [1025,8]; 50 blocks + 2 token-refiner blocks

The main model is not included here. This card does not establish the authorship, license, or redistribution rights of that local fine-tuned checkpoint. The publisher must verify those rights separately.

What was changed

The full H3 AdaLN path consumes the post-SiLU time embedding

S(t) = SiLU(TimeEmbedder(t)) ∈ R^2688

while the curve-pruned checkpoint consumes a shared coordinate table C(t) ∈ R^8. On the 1025-point grid t_j=j/1024, each source AdaLN adapter A ∈ R^(16×2688), B ∈ R^(d_out×16) was converted with an affine least-squares fit:

Y = S @ A.T
X = pinv([C, 1]) @ Y
A8 = X[0:8].T
c = X[8]
diff_b = B @ c

The stored pruned-path update is therefore:

delta(x) = B @ (A8 @ x) + diff_b

The intercept and .diff_b are mandatory. A fit without the constant column loses approximately 94%–99.8% of the original AdaLN Turbo response in the local audit.

Only four FP32 time-embedder tensors were read from the full FL2VA reference checkpoint; its quantized Transformer weights were not used. The 10Eros adaln_t_table raw SHA-256 is ac8727cdec52137c73878d004de5bd2a0e19227e8311e29ab3b68f328310e34e.

Checkpoint integrity

Final LoRA

File: minimax_h3_turbo_4step_10ErosMax_test4_pruned_curveproj1025_exp_v001.safetensors
Size: 794,888,696 bytes
SHA-256: 6c2f38d45dfa3fc282a48de3171b6946a5e6d46e13f832c43b93734f6d12edf5

Structural verification:

• safetensors readback succeeded;
• 259 A/B adapters and 51 .diff_b tensors are present;
• current ComfyUI parsed 259 bypass adapters + 51 regular bias patches;
• 569/569 source keys were consumed;
• all 259 target weight shapes and all 51 bias shapes match the exact 10Eros main model;
• 416 direct adapter tensors are bit-identical to the source;
• all 51 projected-module B tensors are bit-identical to the source;
• no alpha tensors are present, preserving scale 1.0 semantics;
• all three input hashes were unchanged after conversion;
• no .partial file remained.

Detailed evidence:

• evidence/conversion_manifest.json
• evidence/validation.json

ComfyUI usage

Required components

• the exact 10Eros pruned main model and SHA listed above;
• a current MiniMax H3-compatible ComfyUI build;
• a MiniMax H3 Qwen3-VL text encoder;
• minimax_h3_video_vae_fp16.safetensors;
• minimax_h3_audio_vae_fp32.safetensors;
• the minimax-h3-audio-T8 dual-clock nodes used by the validated route.

Suggested layout:

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── 10Eros_Max_h3_fl2va_bf16_test4_pruned.safetensors
    ├── loras/
    │   └── minimax_h3_turbo_4step_10ErosMax_test4_pruned_curveproj1025_exp_v001.safetensors
    ├── text_encoders/
    │   └── qwen3vl_32b_minimax_h3_*.safetensors
    └── vae/
        ├── minimax_h3_video_vae_fp16.safetensors
        └── minimax_h3_audio_vae_fp32.safetensors

Required loading route

1. Load the exact 10Eros model with Load Diffusion Model.
2. Load this LoRA with Load LoRA (Bypass, Model Only) / LoraLoaderBypassModelOnly.
3. Start with LoRA strength 1.0.
4. Use the MiniMax H3 stable dual-clock path with:
   • video steps 4;
   • audio steps 4;
   • video shift 12;
   • audio shift 3;
   • sampler dual_clock_euler;
   • scheduler native_flow.
5. Connect the same packed H3 AV latent to the dual-clock setup and SamplerCustomAdvanced.latent_image.

Do not use a separate custom node that automatically re-injects full-width time conditioning for pruned bases with this file. This LoRA has already converted the 51 AdaLN adapters and includes explicit .diff_b patches; a second pruned-base adaptation can double-apply or change the intended path. The locally validated route is the generic ComfyUI bypass loader plus the T8 dual-clock sampler.

Do not combine this route with an extra Sigma Shift node or an external scheduler. Settings above are the validated baseline, not a guarantee that all prompts will look best at those values.

Validation

Static projection

Metric	Relative error
Stored 1025-point aggregate	9.57255e-5
Stored native-four aggregate	1.58861e-4
Worst native-four module	5.85406e-4

The native-four grid includes video/audio clocks and visual/audio conditioning timestep classes. These are response-preservation metrics, not perceptual-quality scores.

Real four-step AV smoke

ComfyUI: 0.31.0, commit cbbc9dab1f03d0d9a6caa8a8be7d77a7e37e1e44
GPU: NVIDIA GeForce RTX 4060 Ti 16GB
Attention: SageAttention in the live server
Task: T2VA
Resolution: 256 × 256
Frames: 124 at 24 FPS
Steps: 4 video / 4 audio
Sampler/scheduler: dual_clock_euler / native_flow
Shifts: 12 / 3
Seed: 2608104101
Strength: 1.0

The final LoRA and a core208 AdaLN-free ablation both completed real joint H3 generation, dual VAE decoding, and MP4 muxing.

Treatment	Runtime	Video	Decoded audio	Local faster-whisper
curveproj1025	357.84s	124 coherent frames in six-frame screen	stereo 32kHz; 162,816 samples; RMS 0.01879; peak 0.15245; finite; 0% clipping	Today, the rain finally stopped.
core208 ablation	288.58s	124 coherent frames in six-frame screen	stereo 32kHz; 162,816 samples; RMS 0.01222; peak 0.12440; finite; 0% clipping	Today the rain finally stopped

The pair's full-video RGB MAD was 5.615/255, and decoded-audio correlation was 0.719. This proves the projected AdaLN route is active rather than behaviorally identical to dropping the 51 modules; it does not prove that either treatment is perceptually superior.

Limitations

• This is not an official MiniMax, Comfy-Org, LarryVrh, or 10Eros release.
• It is tied to one exact 40GB fine-tuned pruned checkpoint SHA.
• It was not trained, fine-tuned, re-distilled, or calibrated against a broad prompt set.
• The structural projection repairs the pruned AdaLN width mismatch; it cannot guarantee that the fine-tuned main model still follows the original Turbo four-step semantic trajectory.
• Real evaluation covers one prompt, one seed, 256×256, 124 frames, and SageAttention.
• No 20-step no-LoRA quality anchor, four-step no-LoRA negative control, Stock-Attention comparison, high-resolution matrix, held-out prompt/seed matrix, blind listening, identity metric, or trained lip-sync metric has passed.
• The smoke proves intelligible speech and absence of numerical noise/clipping in one case; it does not guarantee universal audio quality.
• More steps, alternative samplers/schedulers, arbitrary strengths, other pruned checkpoints, INT8/FP8 target bases, Ref2VA, and third-party pruned reinjection nodes are unvalidated.
• If held-out four-step quality is unstable or not better than the core208 ablation, no-training static adaptation has reached its evidence boundary and re-distillation/training is required.

Intended use

Intended for controlled local research and A/B evaluation of four-step MiniMax H3 generation on the exact compatible 10Eros main model, subject to all upstream licenses and the acceptable-use policy.

Not intended for claims of universal acceleration quality, arbitrary pruned-base compatibility, identity preservation, deterministic lip sync, medical/forensic use, deceptive impersonation, bypassing safety systems, or any use prohibited by the MiniMax H3 license.

License and attribution

MiniMax H3 and this Model Derivative are governed by the MiniMax H3 Community License Agreement. The upstream license, including territorial restrictions, redistribution conditions, commercial terms, downstream safeguards and acceptable-use policy, takes precedence over this card.

The Turbo source repository declares Apache-2.0 metadata: larryvrh/MiniMax-H3-Turbo-Lora. That does not replace the MiniMax H3 base-model license or any separate rights attached to the local 10Eros fine-tuned checkpoint.

Before distributing this LoRA:

1. include an unmodified copy of the official MiniMax H3 LICENSE;
2. include the provided NOTICE;
3. retain the prominent modification and exact-base notices;
4. verify rights to distribute the 10Eros-dependent derivative;
5. comply with the applicable-territory and downstream-user requirements;
6. verify any commercial authorization requirements directly with MiniMax.

Upstream projects:

• MiniMaxAI/MiniMax-H3
• Comfy-Org/MiniMax-H3
• larryvrh/MiniMax-H3-Turbo-Lora
• MiniMax H3 ComfyUI guide

Modification notice

minimax_h3_turbo_4step_10ErosMax_test4_pruned_curveproj1025_exp_v001.safetensors is an independently modified Model Derivative. It was produced without training by preserving 208 source adapters and converting 51 AdaLN adapters to the exact 10Eros pruned curve basis with an affine 1025-point projection and explicit bias deltas. It is not endorsed by MiniMax, Comfy-Org, LarryVrh, or the author of the 10Eros fine-tuned checkpoint.