MiniMax-H3 INT8 Lean ConvRot — Dynamic Time / Separate QKV

Experimental, patch-required release. These checkpoints do not load in unmodified ComfyUI. Apply the included core patch before loading them. For stock ComfyUI, use the table-based checkpoints in DmitryDB/MiniMax-H3-INT8-Lean-ConvRot.

This repository contains the experimental dynamic-time, physically separate-Q/K/V editions of a quality-oriented mixed-precision INT8 ConvRot conversion of MiniMaxAI/MiniMax-H3. FL2VA and Ref2VA are provided separately because their diffusion-transformer weights are task-specific.

These are community derivatives, not official MiniMax or ComfyOrg checkpoints.

Files

Path	Purpose
`FL2VA/minimax-h3-fl2va-int8-lean-convrot-dynamic-k16-separate-qkv-quality21.safetensors`	FL2VA diffusion transformer, 20.999 GiB
`Ref2VA/minimax-h3-ref2va-int8-lean-convrot-dynamic-k16-separate-qkv-quality21.safetensors`	Ref2VA diffusion transformer, 20.999 GiB
`patches/comfyui-minimax-h3-dynamic-time-separate-qkv.patch`	Required ComfyUI core patch
`reports/validation_dynamic_separate_fl2va.json`	Structural and numerical FL2VA validation
`reports/validation_dynamic_separate_ref2va.json`	Structural and numerical Ref2VA validation
`reports/loadtest_dynamic_separate_fl2va.json`	Full CPU loader result for FL2VA
`reports/loadtest_dynamic_separate_ref2va.json`	Full CPU loader result for Ref2VA
`reports/layer_policy*.json`	Per-layer BF16/INT8 quality policy
`reports/mm_quant_profile*.json`	Row-sampled reconstruction profiles

The MiniMax-H3 text encoder and VAEs are not duplicated here. The two standard VAE files remain available in the stock-compatible repository linked above. A compatible MiniMax-H3 Qwen3-VL text encoder must be installed separately.

What is different

Original FP32 time MLP

The official pruned checkpoints replace the wide time/AdaLN path with a cached table. These checkpoints instead keep all four tensors of the original FP32 time MLP. At each forward pass:

full_t = SiLU(original_time_embedder(t))
coords = full_t @ shared_rank16_basis
AdaLN_i(t) = separate_linear_i(coords)

The rank-16 basis compresses the shared input subspace of the 51 wide AdaLN projections. It does not replace the time MLP with interpolation, and every block plus the final layer keeps an independent FP32 projection.

Physically separate Q, K and V

The source MiniMax Diffusers checkpoints contain separate to_q, to_k, and to_v tensors. Stock ComfyUI normally expects a fused qkv_proj tensor. This release preserves separate projections through loading and forward execution:

50 main transformer attention blocks and two token-refiner blocks;
156 separate Q/K/V projection tensors in total;
no fused qkv_proj weights;
three projection calls per attention block.

The released Q/K/V tensors were checked bit-for-bit against the corresponding contiguous slices used by the stock-compatible quality21 conversion.

Mixed-precision quality21 policy

Both variants use the same quality-oriented policy:

170 of the 200 semantic main transformer matrices use INT8 ConvRot with Hadamard group size 256;
30 high-risk main matrices remain BF16;
token-refiner attention/MLP weights remain BF16;
normalization tensors remain BF16;
patch projections and video/audio output heads retain source FP32;
dynamic time MLP, basis, and all 51 AdaLN projections are FP32.

Because Q, K and V are stored separately, each checkpoint contains 270 physical INT8 ConvRot modules even though the policy represents the same 170 semantic main matrices as the stock-compatible quality21 edition.

Why a patch is required

Unmodified ComfyUI detects the stock MiniMax-H3 layout through fused qkv_proj weights and either adaln_t_table or the standard wide AdaLN construction. It does not construct the combination used here:

independent q_proj, k_proj, and v_proj modules;
original wide FP32 time MLP;
shared dynamic rank-16 coordinates;
51 separate reduced AdaLN projections.

The included patch adds model detection, construction, forward execution, and partial-offload support for this exact layout. No custom node is required after the core patch is applied.

Installation

Apply the patch from the root of your ComfyUI checkout:
```
git apply --check C:\path\to\comfyui-minimax-h3-dynamic-time-separate-qkv.patch
git apply C:\path\to\comfyui-minimax-h3-dynamic-time-separate-qkv.patch
```
The patch was checked against ComfyUI commit 14b05228. If --check fails on a newer checkout, do not force it; port the small changes to the current MiniMax implementation and validate loading again.
Put one or both diffusion checkpoints in:
```
ComfyUI/models/diffusion_models/
```
Install a compatible MiniMax-H3 Qwen3-VL text encoder and the standard MiniMax-H3 visual/audio VAEs.
Use the FL2VA checkpoint for T2VA/I2VA/first-last-frame workflows and the Ref2VA checkpoint for multimodal reference workflows.

Validation performed

The following checks were completed on CPU:

safetensors structure, shapes, dtypes, and all ConvRot descriptors;
1,181 keys and 270 INT8 ConvRot descriptors per checkpoint;
zero fused-QKV keys;
all 156 Q/K/V projection tensors matched their stock quality21 slices bit-for-bit;
all four FP32 time-MLP tensors matched the raw MiniMax-H3 source bit-for-bit;
dynamic AdaLN output relative error on 19 tested timesteps was approximately 2.85e-7 to 3.29e-7;
both files fully loaded through patched ComfyUI as MiniMaxH3Model, with dynamic basis enabled, table mode disabled, and independent Q/K/V modules in the main stack and token refiner.

SHA-256:

495b9d565b17194196646ef39b0c04ac7dde4a8df695b103b3fcbe04b5da00a6  FL2VA/minimax-h3-fl2va-int8-lean-convrot-dynamic-k16-separate-qkv-quality21.safetensors
82d16de12bee9d006c83c80b8843f2bb7e05cd25320724113ca7d128c6be6a04  Ref2VA/minimax-h3-ref2va-int8-lean-convrot-dynamic-k16-separate-qkv-quality21.safetensors

Validation not yet completed

GPU loading and peak-VRAM measurement;
fixed-seed end-to-end FL2VA/Ref2VA generation;
video and audio A/B quality comparison against BF16, official INT8, and the stock table-based quality21 edition;
throughput comparison between fused QKV and three independent projections.

Therefore, the 20.999 GiB file size is a disk-weight target, not a guarantee that every 24 GiB workflow will fit. Runtime memory also depends on activation size, resolution, frame count, attention implementation, and offload settings.

Intended use

This edition is for testing whether retaining the original time encoder and independent Q/K/V projections improves quality or behavior enough to justify a small core patch and extra projection launches. Users who want ordinary ComfyUI compatibility should use the stock table edition instead.

Do not treat CPU structural/numerical validation as proof of perceptual video quality. End-to-end fixed-seed generation remains required.

License

MiniMax H3 is distributed under the MiniMax H3 Community License Agreement. Read LICENSE before using or redistributing these derivatives. The modification summary is provided in NOTICE.

Model tree for DmitryDB/MiniMax-H3-INT8-Lean-ConvRot-Dynamic-Time-Separate-QKV

Base model

MiniMaxAI/MiniMax-H3

Finetuned

(15)

this model

DmitryDB
/

MiniMax-H3-INT8-Lean-ConvRot-Dynamic-Time-Separate-QKV