README.md

metadata

license: other
license_name: minimax-h3-community-license-agreement
license_link: LICENSE
base_model: MiniMaxAI/MiniMax-H3
library_name: comfyui
pipeline_tag: image-text-to-video
tags:
  - minimax-h3
  - comfyui
  - quantization
  - int8
  - w4
  - nvfp4
  - video
  - audio
  - fl2va
  - ref2va

MiniMax-H3 Quants for ComfyUI

Community quantized diffusion-transformer checkpoints for MiniMaxAI/MiniMax-H3. This repository provides the same eight precision profiles for both FL2VA and Ref2VA. All files use the stock ComfyUI fused-QKV and time-table layout: no custom node or core patch is required.

These are community conversions, not official MiniMax or ComfyOrg releases.

1. Choose FL2VA or Ref2VA

• FL2VA — text-to-audio-video, optionally conditioned by a first frame, last frame, or both.
• Ref2VA — reference-to-audio-video using reference images, video, and/or audio.

Download one diffusion checkpoint from the matching column below.

2. Choose a quant

Profile	Direct downloads	Size	What it contains and when to use it
INT8 ConvRot balanced	FL2VA · Ref2VA	20.940 GiB	Recommended for RTX 3090/4090 24 GiB. 170 INT8 + 38 BF16 semantic matrices. Best tested quality/VRAM balance and fully resident in the RTX 4090 loader test.
INT8 ConvRot safe	FL2VA · Ref2VA	20.330 GiB	185 INT8 + 23 BF16 semantic matrices. Choose this on a 24 GiB RTX 30/40 card when the rest of the workflow needs more VRAM.
INT8 ConvRot max	FL2VA · Ref2VA	21.908 GiB	145 INT8 + 63 BF16 semantic matrices. Largest BF16 quality island; recommended for 32 GiB or more. About 0.955 GiB was offloaded in the 24 GiB RTX 4090 loader test.
W8/W4 ConvRot balanced	FL2VA · Ref2VA	13.565 GiB	86 W8 + 114 W4 main matrices; BF16 token refiner. Starting point for 16 GiB RTX 30/40 cards.
W4 ConvRot compact	FL2VA · Ref2VA	10.067 GiB	200 W4 main matrices + 8 INT8 token-refiner matrices. Recommended starting point for 12 GiB cards.
W4 ConvRot offload	FL2VA · Ref2VA	9.708 GiB	All 208 main/refiner matrices use W4. Smallest portable profile; intended for 8 GiB cards with CPU offload.
NVFP4 quality	FL2VA · Ref2VA	13.597 GiB	170 NVFP4 + 30 BF16 main matrices; BF16 token refiner. Recommended for RTX 50/Blackwell 16–24 GiB when NVFP4 support is available.
NVFP4 compact	FL2VA · Ref2VA	10.862 GiB	All 208 main/refiner matrices use block-scaled NVFP4. Smallest Blackwell-specific profile for 8–12 GiB cards.

Short answer

• RTX 4090 24 GiB: start with int8-convrot-balanced; use safe if the workflow needs more activation memory.
• RTX 5090 / Blackwell 16–24 GiB: start with nvfp4-quality for headroom, or INT8 balanced when portability matters.
• RTX 30/40 16 GiB: start with w8w4-convrot-balanced.
• 12 GiB: start with w4-convrot-compact.
• 8 GiB: use w4-convrot-offload and CPU offload.
• 32 GiB or more: int8-convrot-max has the largest BF16 island.

Checkpoint size is not full-workflow peak VRAM. Resolution, frame count, attention backend, text encoder, VAE, and ComfyUI offload settings also matter. RTX 50 recommendations are architecture-based; no RTX 5090 generation run was performed on this machine. NVFP4 here is plain block-scaled NVFP4, not AWQ.

Measured RTX 4090 loader results

FL2VA and Ref2VA were tested independently through stock ComfyUI. Each test also executed a real quantized INT8 projection.

Profile	Loaded weights	Peak reserved	Free after load	Result
int8-convrot-safe	100%	20.424 GiB	2.072 GiB	PASS
int8-convrot-balanced	100%	21.025 GiB	1.471 GiB	PASS
int8-convrot-max	95.6%	21.002 GiB	1.494 GiB	PASS; about 0.955 GiB offloaded

These are loader/kernel measurements, not full prompt-to-decoded-video peaks.

Compatibility

All 16 files in this repository:

• retain all 50 transformer blocks;
• use fused qkv_proj = cat(Q,K,V) tensors expected by stock ComfyUI;
• use a rank-16 FP32, 4,097-point time table;
• retain 51 independent FP32 AdaLN projections;
• load without a custom loader or core patch in the tested ComfyUI revision.

For the original runtime FP32 time MLP and physically separate Q/K/V modules, use the patch-required MiniMax-H3-DynTime-sQKV repository instead.

Advanced: exact INT8 BF16 islands and time/QKV layout

INT8 weights use ConvRot/Hadamard rotation with group size 256, per-row FP32 scales, and deterministic scale search. Norms, conditioning projections, patch projections, output heads, and other small or sensitive tensors retain their source precision.

Profile	BF16 attention-output blocks	BF16 MLP fc2 blocks	Other main semantic matrices
int8-convrot-safe	0, 1, 2, 3, 4, 5, 6, 7, 9, 15, 19, 38, 45, 49	49	INT8 ConvRot
int8-convrot-balanced	0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 17, 19, 20, 27, 38, 43, 44, 45, 46, 47, 49	39, 45, 49	INT8 ConvRot
int8-convrot-max	all blocks 0–49	29, 39, 44, 45, 49	INT8 ConvRot

The eight token-refiner semantic matrices remain BF16 in all three INT8 profiles.

Feature	This stock repository	Patched DynTime s-QKV repository
Attention	One fused projection call	Separate Q, K, and V calls
Original FP32 time_embedder	Absent	Present
adaln_t_table	FP32 [4097,16]	Absent
adaln_curve_basis	Absent	FP32 [2688,16]
adaln_curve_mean	Absent	FP32 [2688]
Per-block AdaLN	51 independent FP32 rank-16 projections	51 independent FP32 rank-16 projections
ComfyUI	Stock	Core patch required

The time table does not remove timestep conditioning. It interpolates a compact representation of the original measured time curve. Maximum measured table interpolation error is below 0.001%; sampled end-to-end AdaLN relative error is approximately 3e-7 to 4e-7 across 19 timesteps.

Validation

Every released checkpoint passed:

1. exact key, shape, dtype, and quantization-inventory checks;
2. sampled reconstruction against its original FL2VA or Ref2VA HF shards;
3. a 19-timestep FP32 AdaLN numerical comparison;
4. complete CPU load as MiniMaxH3Model in clean ComfyUI commit 14b05228;
5. remote byte-size and LFS SHA-256 verification.

Reports are stored under reports/release_matrix/. BF16 samples are checked bit-for-bit. A representative INT8 QKV sample has relative L2 error 0.008814. A prompt-to-decoded-video perceptual A/B score has not yet been measured.

Installation and required components

Place one selected FL2VA or Ref2VA checkpoint in:

ComfyUI/models/diffusion_models/

A complete workflow also needs the separately maintained Qwen3-VL MiniMax-H3 text encoder and these shared VAE files:

File	Role
vae/minimax_h3_video_vae_fp16.safetensors	Video latent encode/decode
vae/minimax_h3_audio_vae_fp32.safetensors	Audio latent encode/decode

No text encoder is included in this repository.

License and attribution

Use is subject to the included MiniMax-H3 community license. The base model is by MiniMax. This community conversion is not endorsed by MiniMax or ComfyOrg.