HyperFlow H3 Flash INT8 (convrot) β FL2VA Γ REF2VA aligned hybrid b25-49
A community hybrid of the two HyperFlow H3 Flash INT8 convrot transformers from BennyDaBall/HyperFlow-H3-Flash-INT8-ComfyUI: the FL2VA file is the base, and source blocks 25β49 take on the REF2VA model's AdaLN modulation.
Unlike a plain tensor swap, the replaced adaln_proj weights are re-fitted to the FL2VA time embedder, so the
blocks reproduce what REF2VA's modulation actually does (see Why aligned below). A plain swap of the same tensors
breaks first/last-frame adherence; this file does not.
Unofficial. Not made, reviewed or endorsed by MiniMax, Video Rebirth, Comfy Org or BennyDaBall.
File
| File | Size | sha256 |
|---|---|---|
minimax_h3_hyperflow_flash_hybrid_aligned_fl2va_ref2va_b25-49_int8_convrot.safetensors |
15.9 GB | e15967787d8836f53fae931a82ae93bbb8973b15bba74a512f0ba0443b50a6b0 |
Same format as the FL2VA source: ComfyUI MiniMax-H3 layout, INT8 convrot, 47 blocks, loads in the
HyperFlow INT8 Loader and the official HyperFlow Loader. Header keeps the FL2VA hyperflow / hyperflow_int8
entries and adds hyperflow_hybrid describing the merge.
What changed
FL2VA HyperFlow Flash INT8 convrot (base: every INT8 code, row scale, low-rank pair,
| embedder, gate and sigma tensor is byte-identical)
+-- source blocks 25-49: adaln_proj.linear.weight / .bias <- REF2VA modulation, re-fitted (BF16)
+-- final_layer.adaln_proj <- FL2VA (unchanged)
- 50 BF16 tensors rewritten (25 blocks Γ weight + bias). Nothing was dequantized or requantized.
- The two source files are pruned differently (FL2VA drops source blocks 6, 21, 22 β 47 blocks; REF2VA drops 9 β
49 blocks), so blocks are matched by their original source block index from each file's
hyperflow_int8.layout: source block b isblocks.{b-3}in FL2VA andblocks.{b-1}in REF2VA.
Why aligned
Each block's adaln_proj reads silu(t_emb), and t_emb comes from the model's own time_embedder (blended with
the HyperFlow endpoint embedder, gate 0.25). FL2VA and REF2VA were distilled separately and their time embeddings
differ a lot (50β140 % relative difference over the 6-step schedule), while the modulation each model actually
produces differs by only 2β5 %. Measured at the six HyperFlow (t, r) pairs:
| per block, relative L2 of the modulation | blocks 25β49 |
|---|---|
| FL2VA vs REF2VA (each with its own embedder) | 1.6β4.1 % |
| plain swap (REF2VA weights fed FL2VA's embedding) vs REF2VA | 8.6β13 % |
| aligned (this file) vs REF2VA | 0.17 % |
The fit: for every replaced block, find Wβ², bβ² with
Wβ² Β· silu(emb_FL2VA(t, r)) + bβ² β W_REF2VA Β· silu(emb_REF2VA(t, r)) + b_REF2VA
by least squares over a dense (t, r) grid plus the 6-step pairs (weighted). Both sides are linear in the augmented
weights, so this is one shared 17Γ17 map applied to each block.
Test renders
RTX 5060 Ti 16 GB, ComfyUI 0.31.0, 1344Γ768, 124 frames, 6 steps, euler, CFG 1, seed 42, patcher: legacy.
First + last frame to video, frames 0 / 60 / 123. Top: FL2VA. Middle: this file. Bottom: plain swap of the same tensors (not published): it ignores the keyframes and re-renders the scene.
This file. Rows 1β2: flf2v dense / Sol sparse attention. Rows 3β4: t2v dense / Sol. Row 5: ref2v with the plain swap. Row 6: ref2v with this file + Sol.
| this file, 16 GB card | total | sampling | s/step | peak VRAM |
|---|---|---|---|---|
| t2v, dense attention | 395 s | 330 s | 51.8 | 15.4 GB |
| flf2v, dense attention | 584 s | 486 s | 65.2 | 15.8 GB |
| t2v, Sol sparse attention | 301 s | 237 s | 22.8 | 14.1 GB |
| flf2v, Sol sparse attention | 284 s | 205 s | 30.6 | 15.7 GB |
| ref2v, Sol sparse attention | 262 s | 190 s | 26.4 | 15.7 GB |
On a 16 GB card about 9.5 GB of the weights stay in system RAM (64 GB here) and stream in each step. Loading the file from a 7200 rpm HDD adds about 80β90 s to the first run.
Use in ComfyUI
Exactly like the FL2VA file of the source repo: put the file in ComfyUI/models/diffusion_models/, install
ComfyUI-HyperFlow-Flash and
comfyui_hyperflow_int8_nodes, and use the source
repo's fl2va workflows (t2v, i2v, flf2v) with this file selected in the loader. It also ran the reference-to-video
workflow in our tests. Text encoder and VAEs: Comfy-Org/MiniMax-H3.
Recommended: HyperFlow Sigmas 6 steps, euler, CFG 1, patcher: legacy.
Tips from testing on an RTX 50 card with a cu128 PyTorch:
- Start ComfyUI with
--enable-triton-backend. Without CUDA 13 PyTorch, comfy_kitchen's CUDA backend is off and INT8 convrot falls back to the slow eager path (~300 s/step here instead of ~52). - Sparse attention (about 2.3Γ faster per step here): the MiniMax-H3 Sol-Attn nodes from
ComfyUI-sol-attn (
MiniMaxH3MemoryEfficientSolAttentionPatchtau 1.3,MiniMaxH3FusedModulation,MiniMaxH3ChunkFeedForward2) work with HyperFlow models, but the pack checks for the exact class nameMiniMaxH3Modeland leaves HyperFlow's subclass unpatched ("expected a MiniMax H3 model"). Changing those checks to also accept subclasses is enough. Sparse attention is an approximation; t2v compositions can differ from dense runs.
License and notices
These files are a Model Derivative of MiniMax-H3, modified from
HyperFlow H3 Flash, distributed under the MiniMax H3 Community License Agreement with no additional or
different terms; LICENSE and NOTICE ship with them as required, and the modification is
recorded in NOTICE and in the file's metadata. The Agreement's territory, commercial-use, use-restriction and
AI-generated-content disclosure terms apply β read it before downloading. Example images in examples/ are frames
of AI-generated video.
Not an official MiniMax product. Powered by MiniMax H3.
Credits: MiniMax (MiniMax H3), Video Rebirth (HyperFlow H3 Flash), BennyDaBall (INT8 convrot solve and ComfyUI conversion), Comfy Org (INT8 convrot kernels), NVIDIA (Sol-Attn).
- Downloads last month
- -

