Explain that keyframes and references can share one generation, and what that needs

README.md

@@ -39,7 +39,8 @@ for reference conditioning. This is **one** checkpoint that serves both, so a de
 
 ## Inference
 
-Both workflows resolve to the weights in this repo, and each loads only its own slot:
+Both workflows resolve to the weights in this repo, and each loads only its own slot — pass a `workflow=`, or the
+components of *both* transformer slots get pulled:
 
 ```python
 import torch
@@ -81,8 +82,27 @@ out = pipe(                       # same repo, reloaded with workflow="fl2va"
 video, audio, rate = out["videos"][0], out["audio"][0], out["sampling_rate"]
 ```
 
-One generation carries **one** shape: the blocks dispatch on their inputs and `references` wins, so passing both
-silently drops the keyframes.
+### Keyframes and references in the same generation
+
+The model itself supports it — MiniMax-H3 denoises one packed sequence, and keyframe conditioning rows and reference
+conditioning rows can both sit in it. ComfyUI exposes this natively: chain its `MiniMaxH3ReferenceToVideo` into
+`MiniMaxH3AddGuide` and one sampler run carries both.
+
+`diffusers` does **not** expose it. Its auto-blocks dispatch either/or — `select_block` checks `references` first and
+returns the `ref2va` branch — so a request carrying both is accepted and the **keyframes are silently dropped**. There
+is no error and no warning; you get a reference-only generation.
+
+Combining them therefore needs one extra piece on top of `diffusers`: a layout step that packs
+`[text | keyframe conditions | reference blocks | targets]` into one sequence, with the keyframe anchors shifted by
+the references' total time span (the references push the target timeline out). Porting
+[ComfyUI's `PackedLayout`](https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/ldm/minimax/model.py) is the
+straightforward way to get it, and the denoising loop needs no change at all, since the conditioning rows are simply
+the leading rows of the sequence.
+
+Measured on this checkpoint with such a layout in place: a combined request (first frame + last frame + an image
+reference) agrees with the real `ref2va` partition at video-latent cosine **0.972** — closer than any reference-only
+request reaches, because keyframes anchoring both ends leave the delta less to carry. Adding an audio reference on top
+moves the video only 0.965 while rewriting the soundtrack to 0.403, i.e. each conditioning does its own job.
 
 ### Few-step generation — keep the turbo LoRA live
 
@@ -114,6 +134,8 @@ ships a `MiniMaxH3PrunedTransformer3DModel` whose AdaLN is 8 wide, where the sto
 5. **Keep the AdaLN affine map.** `adaln_basis` / `adaln_mean` ship as buffers, so LoRAs trained on the released
    2688-wide AdaLN still project onto this 8-wide one.
 
+Verified by re-downloading this repo and generating: bit-identical (`torch.equal` on video and audio latents) to the
+local build it was made from.
 
 ## Measured