Refresh README: link GitHub repo; add L20 directional-hub / RAW=Turbo / refiner-diffuse / rebalance-lever findings; clarify solo probes are diagnostic (attention-mixed slot, not a clean layer)

README.md

@@ -15,43 +15,57 @@ tags:
 
 # Krea 2 Projector Explorations
 
-Interpretability artifacts for Krea 2's text conditioning. Krea 2's text encoder is Qwen3-VL-4B, and the
-DiT doesn't use a single hidden state — it takes **12 selected encoder hidden-state layers**
-`[2,5,8,11,14,17,20,23,26,29,32,35]` (Krea's `select_layers`), concatenates them, and combines them with a
-**learned `Linear(12 → 1)` projector** (`txtfusion.projector`, `[1,12]`). That matrix is the model's own
-per-layer importance — identical in Raw and Turbo:
+Small, Krea-derived interpretability artifacts for Krea 2's text conditioning — the learned layer-mix
+("multilayer feature aggregation") plus single-layer probes. **Full toolkit, methods, figures, and write-up:
+[github.com/fblissjr/krea-explorations](https://github.com/fblissjr/krea-explorations).**
+
+Krea 2's text encoder is a frozen Qwen3-VL-4B; the DiT takes **12 selected encoder hidden-state layers**
+`[2,5,8,11,14,17,20,23,26,29,32,35]` (`select_layers`), combines them with cross-layer attention, then a
+learned `Linear(12 → 1)` projector (`txtfusion.projector`). That matrix is the model's own per-layer
+weighting — **identical in Raw and Turbo** (cosine 1.0):
 
 | layer | L2 | L5 | L8 | L11 | L14 | L17 | L20 | L23 | L26 | L29 | L32 | L35 |
 |-------|----|----|----|----|----|----|----|----|----|----|----|----|
 | w     | -0.05 | -0.16 | +0.37 | +0.50 | +0.71 | +0.39 | +0.40 | **-1.44** | -0.51 | -0.89 | -0.61 | +0.11 |
 
-The deeper layers (L23–L32) carry the most weight. (Terminology: ComfyUI calls these selected layers
-"taps".)
+It combines **contrastively** ("mid plus, deep minus"), not as an average.
 
-## Files
+## What we measured
 
-- `krea2_projector_original_weights.safetensors` — the learned projector weights above, as a reference.
-- `solo/projector_solo_bNN_Lxx.safetensors` — 12 single-layer probes. Each isolates one selected
-  hidden-state layer (zeroing the other 11) so you can see what that individual layer contributes to the
-  image.
+These are characterizations of an open model's learned behavior (not architecture — the architecture is
+public); most are low-effort to reproduce. Full method + confidence levels in the GitHub repo.
 
-Each `solo/` file is a `diffusion_model.txtfusion.projector.diff` patch (one `[1,12]` tensor, ~300 bytes),
-so it loads as a standard ComfyUI LoRA — no custom node needed.
+- **L20 is a learned *directional* attention hub.** In the cross-layer attention, ~91–95% of content tokens
+  route to layer 20 — content-driven (not a padding artifact) and a *directional* effect (not a magnitude
+  sink). Holds across 5 prompts and on **both Raw and Turbo**. The token-side "refiner" blocks, by contrast,
+  are diffuse (no hub).
+- **The projector-rebalance lever is a detail/intensity knob, not an attribute gate.** Benign attributes
+  (expression, "wet", blush) come through the aggregation and render with or without rebalancing; boosting
+  the deep layers mainly shifts detail / contrast / intensity — consistent with the deep layers carrying
+  fine detail.
 
-## Usage (ComfyUI)
+Per-layer reweighting of Krea 2's conditioning was introduced by
+[nova452/ComfyUI-ConditioningKrea2Rebalance](https://github.com/nova452/ComfyUI-ConditioningKrea2Rebalance)
+and refined by [huwhitememes/comfyui-krea2-conditioning](https://github.com/huwhitememes/comfyui-krea2-conditioning).
 
-Put a `solo/` file in `models/loras/`, add a **LoraLoaderModelOnly** node between your model loader and the
-sampler, and run it at **strength 1**. The image is then driven by a single selected layer's contribution.
-Sweep across the 12 to compare. Works with Krea 2 Raw or Turbo (the projector is identical).
+## Files
+
+- `krea2_projector_original_weights.safetensors` — a **reference copy** of the 12 learned projector weights
+  above (the `[1,12]` tensor itself). Read-only reference, not a LoRA to apply.
+- `solo/projector_solo_bNN_Lxx.safetensors` — 12 **diagnostic probes**. Each is a projector `.diff` that, at
+  strength 1, keeps one of the projector's 12 inputs and zeroes the other 11, so the DiT conditions on a
+  single slot — useful to *see* what that slot contributes (deep slots render coherent images, shallow are
+  noise, L14 carries text/structure, L35 alone is unusable).
 
-## How they were made
+**Important:** the projector's 12 inputs are the **attention-mixed slots** (output of the 2 layerwise blocks),
+not pristine encoder layers — and because the cross-layer attention routes through L20, every slot already
+carries L20 content. So a "solo Lx" isolates the slot *indexed by* layer x, not a clean layer x. These are
+**interpretability probes, not generation LoRAs** (keeping one input by design gives a partial/degraded image).
 
-A `solo` patch keeps one selected layer and zeros the rest: `diff = original·(gain − 1)` with `gain = 1`
-for the kept layer and `0` elsewhere, so at strength 1 the projector reduces to that single layer.
-Deterministic edits of the model's existing weights — no training.
+Each `solo/` file is a `diffusion_model.txtfusion.projector.diff` patch (one `[1,12]` tensor, ~300 bytes),
+loadable via the stock `LoraLoaderModelOnly` — no custom node. (ComfyUI calls the selected layers "taps".)
 
 ## License
 
-Derivatives of Krea 2 weights, released under the **Krea 2 Community License**
-([license](https://huggingface.co/krea/Krea-2-Raw/blob/main/LICENSE.pdf) · base model
-[krea/Krea-2-Raw](https://huggingface.co/krea/Krea-2-Raw)).
+These artifacts derive from Krea 2 and are covered by the **Krea 2 Community License** (see the base model
+linked above).