Update README.md

README.md

@@ -12,12 +12,10 @@ tags:
 ---
 
 
-This is a cross attention transfer from Eros 1.5 to stylize and alter minimax-H3 based on an LTX model in it's first experimental version. That's a cross model transfer but only at gate head level and certain block attention, 32 ltx heads mapped to a region on H3's 56 and adjusting attn_self from LTX attn_1. After a lot of tinkering the peaks and archs were aligned allowing larger motion style changes and deviation/influence without directly touching or changing the weights, adjusting any safety features, or introducing any foreign data into the model's weights, all in-line with the license. It is not specialized or modified to produce anything other than base material, but will guide itself differently based on a different trajectory. Also worth noting; I'm in Beijing, China-- using in-region.
+So far, this release is an experiment on MiniMax H3 base model that combines learned patterns from two other video diffusion models — LTX 2.3 and Wan 2.2 — into H3's transformer architecture. 
 
-这是从 Eros 1.5 进行的交叉注意力迁移，用来对基于 LTX 模型的 MiniMax-H3 进行风格化与调整，目前是第一版实验。
-本质上属于跨模型迁移，但只作用于门控头（gate head）级别以及部分 block 的注意力层：把 LTX 的 32 个注意力头映射到 H3 共 56 个头中的某个区域，并调整来自 LTX attn_1 的自注意力（attn_self）。
-经过大量调试后，峰值与拱形结构已对齐，因此可以在更大范围内改变运动风格并产生偏差/影响，同时完全不直接修改权重、不触碰任何安全机制、也不向模型权重中引入任何外来数据，全程符合许可证要求。
-它并没有被专门化或改造成只生成某种特定内容，仍然只输出基础素材，只是会沿着不同的轨迹进行自我引导。
-另外补充一点：我目前人在北京，使用的是国内区域。
+H3's transformer is organized into 50 blocks, each containing self-attention and feed-forward layers that process video, audio, and text together in a single unified stream. This work modified those attention and feed-forward layers in specific blocks, blending in the corresponding weights from LTX 2.3 and Wan 2.2. Nothing outside those layers was touched — H3's input processing, output projections, timestep modulation, and all the machinery that decides how to route different modalities remained fully intact and original.
 
-The model itself is very uncensored and can generate very explicit generations and fairly grotesque gore generations, not sure exactly what they distinguish as out of their guardrails in the licesne. The model has no signatures of safety filters in it based on block analysis and maps I did, there are none to remove or change that I see. All this does is emphasize and express certain themes or prompts more easily in it's own way of motions v.s. the standard model. 
+The LTX pass was applied first to H3's front blocks, followed by a Wan pass targeting H3's middle-to-back blocks. Each pass modified attention weights directly; the second pass also modified the feed-forward layers in a broader block range. It successfully shows that two completey different architecture video model donors contribute different character into H3 that is recognizable. MiniMax's own architecture documentation notes that H3's attention and feed-forward layers do not contain modality-specific structure — the parts of the model that handle "this is audio vs video vs text" live elsewhere in the model. That means the attention and feed-forward weights are the architecturally-appropriate location to transfer character from other trained models without disrupting how H3 handles each modality.
+
+The modifications use orthogonal projection math — new weight patterns are added into H3 primarily in directions the base model wasn't already using. This lets the transferred character supplement H3 rather than override it. Combined with careful control over which blocks are affected, this preserves H3's core capability to produce coherent video with audio while shifting its aesthetic and motion character toward the donor models without modifying the model's actual blocks weight, safety features, or guardrails. The standard H3 community license applies, and also includes the LTX one as well since technically this is now an LTX model to a certain extent.