Instructions to use austinit/TaoMate-H3-FL2VA-ComfyUI with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Inference
- Notebooks
- Google Colab
- Kaggle
TaoMate-H3-FL2VA 3-Step LoRA for ComfyUI
ComfyUI-compatible conversions of the official
TaoLiveAIGC/TaoMate-H3-FL2VA
3-step LoRA adapter for MiniMax H3 FL2VA. Two variants are provided: the full
rank-128 conversion and a much smaller dynamic-SVD version with average rank 19.
这是官方 TaoMate-H3-FL2VA 三步 LoRA 的 ComfyUI 格式转换,仓库同时提供完整 rank-128 版本与动态 SVD 平均 rank-19 版本。
Modified files: both
.safetensorsfiles in this repository are modified derivatives of the official adapter. Key names, dtypes, alpha tensors and, for the compact variant, matrix ranks were changed for ComfyUI compatibility.
Files
| File | Rank | Size | SHA-256 |
|---|---|---|---|
TaoMate-H3-FL2VA-ComfyUI.safetensors |
128 (fixed) | 1,240,539,224 bytes | df73d895f0ba112fb4b5adeb4acf6651de85edb8190f35d66da4562df1f1c70b |
minimax_h3_taomate_fl2va_3step_avg_rank_19_bf16.safetensors |
19.365 average (dynamic 2–100) | 176,723,664 bytes | e6f86322a0158dcf804b106e784a9ec1a078b1522c161143627364edc08a8bd9 |
Full rank-128 version
- 416 LoRA matrices stored as BF16.
- 208 FP32 alpha scalars, all set to 128.
- 624 tensors forming 208 complete LoRA patches.
- Every converted matrix was checked element-by-element against the official FP32 source cast to BF16.
Conversion mapping:
- Add the
diffusion_model.prefix. - Rename
.lora_a/.lora_bto.lora_A.weight/.lora_B.weight. - Cast LoRA matrices from FP32 to BF16.
- Add one FP32
.alphascalar with value 128 per module.
Dynamic average-rank-19 version
This is not fixed rank 19. Each module is independently approximated using a
truncated SVD of its effective update matrix B @ A. Dynamic ranks are selected
with the KJNodes/kohya-style sv_fro=0.99 rule. The result has:
- 208 modules with ranks ranging from 2 to 100.
- Average rank: 19.365385.
- Per-module
alpha = rank, keeping the effective LoRA scale at 1. - 624 tensors, all stored as BF16.
- Approximately 169 MiB on disk.
This compact file is byte-identical to Kijai's published
minimax_h3_taomate_fl2va_3step_avg_rank_19_bf16.safetensors.
It originates from the same official TaoMate-H3-FL2VA step-300, rank-128 adapter.
Credit to Kijai and the KJNodes/kohya LoRA-resize implementation for this
dynamic-rank result.
ComfyUI usage
Download either file into ComfyUI/models/loras/, then load it with the standard
LoraLoader or LoraLoaderModelOnly node. Use a MiniMax H3 FL2VA base model
and ComfyUI nodes that support first/last-frame FL2VA conditioning.
Suggested starting settings from the adapter metadata and the 3-step student:
| Setting | Value |
|---|---|
| LoRA strength | 1.0 |
| Steps | 3 |
| CFG | 1.0 |
| Sampler | Euler |
| Sigma schedule | 1.0, 0.961165, 0.853333, 0.0 |
| Audio shift | 3 |
The sigma values are the whole-clip approximation recorded in the compact file's metadata (50-point shift-12 schedule at indices 0/16/33/49). Workflow/node implementations may expose these controls differently.
The LoRA conversion alone does not turn a regular H3/T2AV base into FL2VA. First/last-frame conditioning and streaming AV-KV behavior require the proper FL2VA base model and compatible runtime/nodes.
Download
# Full rank-128 version
hf download austinit/TaoMate-H3-FL2VA-ComfyUI \
TaoMate-H3-FL2VA-ComfyUI.safetensors \
--local-dir ComfyUI/models/loras
# Compact dynamic-rank version
hf download austinit/TaoMate-H3-FL2VA-ComfyUI \
minimax_h3_taomate_fl2va_3step_avg_rank_19_bf16.safetensors \
--local-dir ComfyUI/models/loras
Validation
Both files were structurally validated with safetensors:
- exactly 208 complete LoRA modules and 624 tensors;
- matching A/B ranks and alpha values;
- expected dtype for every tensor;
- no NaN or Inf values;
- recorded SHA-256 checksums verified after conversion.
Full end-to-end video inference was not performed as part of this conversion.
Credits
- Original adapter and FL2VA work: Alibaba TaoLive AIGC Team
- Base model: MiniMaxAI/MiniMax-H3
- Dynamic-rank conversion/reference: Kijai/MiniMax-H3-experimental
- Earlier H3 ComfyUI naming reference: CZMartin22/TaoMate-H3-3step-ComfyUI
License and territory restriction
These files are distributed under the MiniMax H3 Community License Agreement.
See LICENSE and NOTICE before downloading or using them.
The license defines the Applicable Territory as worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America. Use, reproduction, modification, distribution, and display outside the Applicable Territory are not authorized by that agreement. Additional acceptable- use, downstream-user, commercial, attribution, and safeguards requirements also apply. This summary is not a substitute for the full license.
Powered by MiniMax H3.
- Downloads last month
- 738
Model tree for austinit/TaoMate-H3-FL2VA-ComfyUI
Base model
MiniMaxAI/MiniMax-H3