Update README.md

README.md

@@ -26,7 +26,19 @@ The modifications use orthogonal projection math — new weight patterns are add
 
 Combined with careful control over which blocks are affected — and, for Krea specifically, careful control over which attention head positions receive the injection to avoid disturbing the head slots most involved in audio processing — this preserves H3's core capability to produce coherent video with audio while shifting its aesthetic and motion character toward the donor models. The model's actual block weights, safety features, and guardrails remain unmodified in the architectural sense — every modification is a controlled additive perturbation of existing weights, not a replacement of the model's own structure.
 
-In the final beta1 configuration, the other model graft donors were given surgical targets. The H3 model's attn triplets were unfused resulting in a more normal attn strucure, although not usable in inference. Mainly in model blocks blocks 0 and 1 tokenization, which control essentially the entire prompting direction and overall output structure. Tokenization attn_q was given to Krea to alter the model to draw like Krea does but at token level. attn_k was isolated. This is a very important finding for training and people merging into the model, in my opinion attn_k should be kept pristine on all blocks and elements. Any alteration, even slight, causes the audio modality of the model to wither along with many of the other perfectly trained small details and relations present from the model's reinforcement learning which makes sense as the H3 model has no control gates, instead relying on the delicately dialed attn_k to control amplification. That should not be trained on to, but it is, in all training pipelines right now. The model's triplets are easily unfused, isolated, and refused with proper code. attn_v however carries a large amount of output chracter, almost all of it, an that was given to the LTX2.3 Eros model to expand with a linear magnitude graft. The triplets were then refused, and Wan2.2 was give blocks 0,1 MLP fc1, which by dimension is almost a perfect slot for Wan2.2. Overall they combine to inject a substantial but very subtle influence into the model, bring the small details to certain motions that I wanted to see carried over, and alter t2v styling. They do all of this without degrading anything except audio quality slightly, and they remove no prompting or behavior capability from the base model like most training and merging does. Audio loss however is essentially unavoidable with how the model's audio is unified. The unified audio I now view as a flaw in the model's open source usability. 
+In the final beta1 configuration, the H3 model received surgical grafts from other donors targeted at the token_refiner blocks 0 and 1 — the two-block text refinement stack that processes Gemma text embeddings before they condition main-block generation. Because token refinement shapes what the main blocks receive as text conditioning, changes here influence the entire prompting direction and output structure.
+
+The H3 model's fused Q/K/V projections were unfused into separate q_proj, k_proj, v_proj tensors — necessary because the fused format doesn't allow per-band surgery, but does require refusing them back into the fused qkv_proj format for ComfyUI inference.
+
+attn_q was given to Krea to shift the token refinement's "what to attend to" direction toward Krea's learned patterns, altering styling at the token level.
+
+attn_k was isolated. This is worth naming as an empirical finding for anyone training or merging into H3: attn_k appears to carry critical calibration for the model's amplification behavior. Even slight perturbation degrades audio quality along with other finely-trained details from the model's reinforcement learning. Since H3 doesn't have explicit control gates (unlike LTX which uses to_gate_logits), it appears to rely on precisely-tuned attn_k to modulate attention strength. Standard training pipelines that hit K without knowing this may be silently degrading the model. Isolating K requires unfusing the qkv triplet, targeting only Q and V, and refusing — which the beta1 toolchain supports.
+
+attn_v was given to LTX-2.3 (Eros lineage) via linear-mag mode graft. V carries a substantial portion of the output character, so this was where LTX Eros content flows into H3's text refinement.
+
+The triplets were refused back to fused qkv_proj format, and then Wan 2.2's main-block MLP fc1 was grafted into blocks 0 and 1 as a supplementary contribution — Wan 2.2's dimensions align nearly perfectly with H3's token_refiner MLP inner (96% coverage), making it structurally well-suited.
+
+Together these grafts inject a substantial but subtle influence: they bring specific motion details across, adjust t2v styling, and do this without removing prompting or behavior capability from the base model in the way many training and merging approaches do. Some audio quality degradation is present but limited. Given H3's unified audio-video attention architecture, meaningful modification of the video path is difficult to separate from audio impact — the unification is genuinely difficult to work around at the merge/graft level.
 
 
 The standard H3 community license applies. Because this release now carries transferred character from LTX 2.3, Wan 2.2, and Krea 2, the community licenses for those source models apply as well to the portions of character that came from each.