Model card and NOTICE: the fl2va tower, link the loader

NOTICE

@@ -7,15 +7,16 @@ under the same Agreement. This repository redistributes them under that Agreemen
 the Agreement is included as LICENSE.
 
 Modifications from Video Rebirth's release (Section III.2), in
-diffusion_models/minimax_h3_ref2va_hyperflow_flash_int8_convrot_comfy.safetensors:
-- Converted from ref2va_int8_convrot/ (diffusers layout) to ComfyUI's int8_tensorwise + convrot layout. The int8
-  values and per-row scales are copied unchanged; no weights were retrained or requantized.
-- Tensors are renamed to ComfyUI's MiniMax H3 names, q/k/v rows are concatenated, and blocks after the dropped block
+diffusion_models/minimax_h3_fl2va_hyperflow_flash_int8_convrot_comfy.safetensors (from fl2va_int8_convrot/) and
+diffusion_models/minimax_h3_ref2va_hyperflow_flash_int8_convrot_comfy.safetensors (from ref2va_int8_convrot/):
+- Converted from the diffusers layout to ComfyUI's int8_tensorwise + convrot layout. The int8 values and per-row
+  scales are copied unchanged; no weights were retrained or requantized.
+- Tensors are renamed to ComfyUI's MiniMax H3 names, q/k/v rows are concatenated, and blocks after the dropped blocks
   are renumbered.
 - The two halves of each SwiGLU fc1 projection are swapped to put the gate first.
 - Low-rank pairs are kept in bf16 with their rank zero-padded to a multiple of 256, which leaves their product
   unchanged.
-- Every converted tensor was checked against Video Rebirth's ref2va Q8_0 GGUF of the same weights.
+- Every converted tensor was checked against Video Rebirth's Q8_0 GGUF of the same weights.
 
 This repository is not an official MiniMax or Video Rebirth product and is not affiliated with or endorsed by either.
 Powered by MiniMax H3.

README.md

@@ -29,13 +29,19 @@ These files are Model Derivatives of MiniMax H3. Both the original model's licen
 
 ## HyperFlow H3 Flash
 
-`minimax_h3_ref2va_hyperflow_flash_int8_convrot_comfy.safetensors` is Video Rebirth's ref2va `int8_convrot` build in ComfyUI's int8 layout. It is a 14B tower with one block dropped and 19 blocks factored into low-rank pairs, and it samples in 6 or 8 steps without PDD. The int8 weights are copied unchanged from Video Rebirth's file.
+Video Rebirth's two HyperFlow towers in ComfyUI's int8 layout. Both sample in 6 or 8 steps without PDD, and the int8 weights are copied unchanged from Video Rebirth's files.
 
-It needs a loader that knows HyperFlow's endpoint time embedder and low-rank blocks; ComfyUI's stock H3 loader won't load it. Use it with the stock MiniMax H3 text encoder and VAEs from [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3).
+| File | Tower | Use it for |
+|---|---|---|
+| `minimax_h3_fl2va_hyperflow_flash_int8_convrot_comfy.safetensors` (16.3 GB) | fl2va: 3 blocks dropped, 102 linears low-rank | Text to video, and image to video from a first frame or first and last frames (`MiniMaxH3ImageToVideo`) |
+| `minimax_h3_ref2va_hyperflow_flash_int8_convrot_comfy.safetensors` (16.8 GB) | ref2va: 1 block dropped, 114 linears low-rank | Video from reference images and audio (`MiniMaxH3ReferenceToVideo`) |
+
+They need [ComfyUI-Studio-HyperFlow](https://github.com/NoDevOrg/ComfyUI-Studio-HyperFlow)'s `HyperFlowInt8Loader` and Video Rebirth's [ComfyUI-HyperFlow-Flash](https://github.com/Video-Rebirth/ComfyUI-HyperFlow-Flash) for the schedule; ComfyUI's stock H3 loader won't load them. Use them with the stock MiniMax H3 text encoder and VAEs from [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3).
 
 ```
 📂 ComfyUI/
 ├── 📂 models/
 │   ├── 📂 diffusion_models/
+│   │   ├── minimax_h3_fl2va_hyperflow_flash_int8_convrot_comfy.safetensors
 │   │   └── minimax_h3_ref2va_hyperflow_flash_int8_convrot_comfy.safetensors
 ```