MiniMax-H3-fl2va-ref2va-b25-49-r1024-r0

MiniMax-H3 hybrid (fl2va × ref2va) — fp8_scaled fused model collection

This is a Model Derivative built by fusing MiniMax AI's official MiniMax-H3 checkpoints.

What this is

The two official checkpoints are merged into a single self-contained fused model that carries both pipelines in one file:

• fl2va (base): first/last-frame keyframes / image quality / audio quality
• ref2va (capability source): reference-image / reference-video / reference-audio driving

The fusion method is neither training nor LoRA — it is tensor-level weight fusion. Three headline tiers are provided, ordered by reference-injection strength (b25-49 → r1024 → r0; the further right, the stronger the reference capability and the closer to ref2va; the further left, the more fl2va's native orientation is preserved). There are also AdaLN block-range variants b15 / b20 / b30 (see below; only the starting block differs), ordered along the "preserve fl2va quality + light reference" axis:

All three tiers share the same key layout and quantization format, and can be swapped/loaded at any time. Use b25-49 for "maximum fl2va quality + a little reference"; use r1024 for "balanced reference and quality"; use r0 for "maximum reference, upper-bound comparison".

AdaLN block-range variants (b15 / b20 / b30)

Same origin and method as b25-49 (AdaLN modulation-projection block-range replacement, trunk byte-identical); the only difference is the starting block, which determines how much of the ref2va reference injection is covered:

Rule: the smaller the starting block → the more blocks injected from ref2va → the stronger the reference capability, but the more it deviates from fl2va's native quality orientation. This family (b15 / b20 / b30 / b25-49) only touches adaln; the trunk is byte-identical, so overall quality loss is controllable — it belongs to the "light reference enhancement" line. Compared with r1024 / r0 (full-tensor Delta-Fused), the reference injection is shallower but safer (it does not touch the trunk structure).

Tier suggestion: use b30 for "maximize fl2va quality + minimal reference"; use b25-49 / b20 for "a bit more reference"; use b15 for "reference-first, slight quality trade-off". They share the same key layout as the three headline tiers and can be swapped directly for comparison.

Reference measurements (weight-level / proxy metrics, for tier selection)

• fl2va and ref2va are 97%+ identical at the bit level; the real differences concentrate in each block's adaln_proj (AdaLN modulation projection) and the trunk difference.
• Tier essence (measured on a sample of 10 trunk weights + 5 adaln relative norms):
  • b25-49: trunk has zero difference from fl2va; adaln only takes the latter half (25–49) from ref2va → shallowest reference injection
  • r1024: trunk injects roughly half the main body of the difference toward ref2va; adaln fully taken from ref2va → reference capability clearly enhanced
  • r0: trunk has zero difference from ref2va (full injection) → equivalent to reference upper bound
• Community xmarre proxy comparison on similar tiers (video-latent, cosine; larger = closer to true ref2va reference behavior): block-range adaln tier ≈0.505, rank-1024 delta tier ≈0.691 (video-ref dimension); photoreal dimension 0.678 → 0.875. Final look-and-feel should be judged by actual generation comparisons.

Format

• fp8_scaled (ComfyUI comfy_quant format: weights F8_E4M3 + per-tensor weight_scale, scale = max(|W|)/448)
• Each tier is ~19.5 GB; at inference the weights occupy only ~10 GB in VRAM
• Unquantized tensors (adaln_proj / RMSNorm / bias, etc.) keep their original precision

Supported hardware

• Tested usable, stable, and VRAM-safe on NVIDIA V100 (Volta, SM 7.0).
• V100 has no FP8 hardware; fp8 weights are 1 byte/element in VRAM (half the size), de-quantized to FP16 at compute time and run through native FP16 Tensor Cores.
• Best results with the ComfyUI-MiniMaxH3-SolAttn-V100 patch node (FP16Safe + sparse attention).
• Ada/Blackwell (40/50 series) natively support FP8 matrix multiply, also applicable and faster.

Usage

Load with ComfyUI's ordinary Load Diffusion Model node (self-contained checkpoint, no custom loader needed). It is routed as ref2va; both the first/last-frame keyframe and reference-image pipelines are available simultaneously. The three tiers can be loaded side-by-side for comparison; pick a tier by use case.

Source models

• MiniMaxAI/MiniMax-H3 (the pruned_fp8_scaled versions of fl2va + ref2va)
• Encoder dependency: Qwen3-VL-32B (Apache 2.0)

⚠️ License & compliance (important)

This repository is a derivative work of MiniMax-H3, governed by the MiniMax H3 Community License Agreement (not MIT). Redistribution must satisfy:

1. Include the full license text (see LICENSE in this repo);
2. Include the NOTICE file (already included);
3. Modified files must carry a prominent modification statement (this README is the statement):
   • b25-49: using fl2va as base, blocks 25–49 adaln_proj.linear.* taken from ref2va;
   • b15 / b20 / b30: same method as b25-49, only the starting block is 15 / 20 / 30 respectively (i.e. blocks 15–49 / 20–49 / 30–49 adaln_proj.linear.* taken from ref2va);
   • r1024: using fl2va as base, apply (ref2va − fl2va) difference to all model tensors, trunk difference injected at low-rank rank-1024, RMSNorm/bias/adaln_proj etc. precisely replaced with ref2va;
   • r0: using fl2va as base, apply (ref2va − fl2va) difference to all model tensors at full rank (no truncation, equivalent to ref2va weight orientation);
4. Commercial products with annual revenue over 20 million USD require separate written authorization from MiniMax ([email protected]);
5. Commercial UIs must label "MiniMax H3".

Territorial restriction: applies globally except the European Union, United Kingdom, South Korea, and United States. Public repos are globally accessible; for strict compliance set the repo to private / gated to avoid distribution to excluded territories.

---

中文版（Chinese Version）

MiniMax-H3 hybrid (fl2va × ref2va) — fp8_scaled 融合模型集

这是一个衍生模型（Model Derivative），基于 MiniMax AI 官方发布的 MiniMax-H3 融合而成。

这是什么

把官方两个 checkpoint 拼成一个自包含融合模型，单文件同时具备两条通路：

• fl2va（基座）：首尾帧关键帧 / 画质 / 音质
• ref2va（能力来源）：参考图 / 参考视频 / 参考音频驱动

融合方法不是训练、不是 LoRA，而是张量级的权重融合。主推 3 个档位（按参考注入强度排列 b25-49 → r1024 → r0），另有 AdaLN 块区间变体 b15 / b20 / b30（见下文，仅改起始 block），按「保留 fl2va 画质 + 轻量参考」取向排列：

三档共用同一套 key 布局与量化格式，可随时互换加载。想要「保留最多 fl2va 画质 + 少量参考」用 b25-49；「参考与画质平衡」用 r1024；「参考拉满、对照上限」用 r0。

AdaLN 块区间变体（b15 / b20 / b30）

与 b25-49 同源同法（AdaLN 调制投影块区间替换、trunk 逐字节不动），区别仅在起始 block 不同，从而决定 ref2va 参考注入的覆盖范围：

规律：起始 block 越小 → 注入 ref2va 的 block 越多 → 参考能力越强，但越偏离 fl2va 原生画质取向。这一家族（b15 / b20 / b30 / b25-49）都只动 adaln、trunk 逐字节未动，画质损失整体可控，属于「轻量参考增强」一条线；与 r1024 / r0（全张量 Delta-Fused）相比参考注入更浅、但更安全（不触碰主干结构）。

选档建议：想要「fl2va 画质最大化 + 少量参考」用 b30；「参考再强一点」用 b25-49 / b20；「参考优先、画质略让」用 b15。它们与三档主推模型共用同一 key 布局，可直接互换对比。

实测参考（权重层/代理指标，供选档参考）

• fl2va 与 ref2va 97%+ 权重逐比特相同，真正差异集中在每 block 的 adaln_proj（AdaLN 调制投影）与 trunk 差分。
• 档位本质（抽样 10 个 trunk 权重 + 5 个 adaln 的相对范数实测）：
  • b25-49：trunk 与 fl2va 零差异；adaln 仅后半(25-49)取自 ref2va → 参考注入最浅
  • r1024：trunk 向 ref2va 方向注入约一半差分主体；adaln 全部取自 ref2va → 参考能力明显增强
  • r0：trunk 与 ref2va 零差异（全量注入）→ 等价参考上限
• 社区 xmarre 对类似档位的 video-latent 代理对照（cosine，越大越接近真 ref2va 的参考行为）：块区间 adaln 档 ≈0.505、rank-1024 delta 档 ≈0.691（video-ref 维度）；photoreal 维度 0.678 → 0.875。最终观感请以实际生成对比为准。

格式

• fp8_scaled（ComfyUI comfy_quant 格式：权重 F8_E4M3 + per-tensor weight_scale，scale = max(|W|)/448）
• 每档文件约 19.5 GB，推理时权重在显存中只占约 10 GB
• 未量化张量（adaln_proj / RMSNorm / bias 等）保持原精度

适用硬件

• 在 NVIDIA V100（Volta, SM 7.0） 上实测可用、稳定、不爆显存。
• V100 无 FP8 硬件，fp8 权重在显存中是 1 字节/元素（省一半），计算时反量化到 FP16 走原生 FP16 Tensor Core。
• 配合 ComfyUI-MiniMaxH3-SolAttn-V100 补丁节点（FP16Safe + 稀疏注意力）效果最佳。
• Ada/Blackwell（40/50 系）原生支持 FP8 矩阵乘，同样适用且更快。

使用方法

用 ComfyUI 普通的 Load Diffusion Model 节点加载即可（自包含 checkpoint，无需自定义 loader）。识别为 ref2va 路由，首尾帧关键帧与参考图两条通路同时可用。三档可加载后并排对比，按用途选档。

源模型

• MiniMaxAI/MiniMax-H3（fl2va + ref2va 的 pruned_fp8_scaled 版本）
• 编码器依赖 Qwen3-VL-32B（Apache 2.0）

⚠️ 许可证与合规（重要）

本仓库是 MiniMax-H3 的衍生作品，受 MiniMax H3 Community License Agreement 约束（不是 MIT）。再分发需满足：

1. 附带完整协议文本（见本仓库 LICENSE）；
2. 附带 NOTICE 文件（已包含）；
3. 修改文件须有醒目修改声明（本说明即声明）：
   • b25-49：以 fl2va 为基座，blocks 25–49 的 adaln_proj.linear.* 取自 ref2va；
   • b15 / b20 / b30：与 b25-49 同法，仅起始 block 分别为 15 / 20 / 30（即 blocks 15–49 / 20–49 / 30–49 的 adaln_proj.linear.* 取自 ref2va）；
   • r1024：以 fl2va 为基座，全模型张量施加 (ref2va − fl2va) 差分，trunk 差分以 rank-1024 低秩注入，RMSNorm/bias/adaln_proj 等精确替换为 ref2va；
   • r0：以 fl2va 为基座，全模型张量全量施加 (ref2va − fl2va) 差分（无截断，等价 ref2va 权重取向）；
4. 商业产品年营收超 2000 万美元需 MiniMax 单独书面授权（[email protected]）；
5. 商业 UI 须标注 "MiniMax H3"。

地域限制：适用地域为全球除欧盟、英国、韩国、美国。公开仓库全球可访问，若需严格合规请设为 private / gated，避免向 excluded territories 分发。