TaoMate-H3-FL2VA is Alibaba TaoLive AIGC's new three-step adapter and streaming implementation for MiniMax H3 with first/last-frame control. The October 15 release is distinct from the earlier TaoMate-H3 text-to-audio-video checkpoint. Downloading either adapter does not replace the base model, and loading a converted LoRA in ComfyUI does not install the streaming runtime.
This guide follows the official FL2VA model card, runtime repository, and conversion authors' documentation, checked October 16, 2026. It is a documentation-based setup guide; we have not generated a comparison or verified consumer-GPU performance. H3Vid does not currently execute TaoMate.
Choose the checkpoint before the file size
The Robert1212star conversion repository contains two separate source checkpoints. Its old taomate_h3_3step_comfy.safetensors belongs to the original release. The new file is TaoMate-H3-FL2VA-3Step-ComfyUI-BF16.safetensors. The latter is approximately 1.24 GB and converts the new adapter from FP32 to BF16 while preserving its rank-128 layout, including token-refiner tensors. Reduced precision changes the stored values; it is not lossless compression.
Choose by source identity, then precision and loader compatibility. An old file remaining in the same repository does not become the new FL2VA release. Likewise, a smaller filename containing “TaoMate” does not establish which checkpoint it derives from or reproduce the official memory policy. Keep the download URL and revision with your workflow so a later comparison remains interpretable.
What ordinary ComfyUI loading provides
For the named BF16 conversion, download the selected file into ComfyUI/models/loras/, refresh the model list, and select it in a MiniMax H3-compatible LoRA loader. Use a compatible H3 base and conditioning graph. Inspect the console for unused keys and tensor-shape errors before trusting that the whole adapter loaded.
The conversion author checked tensor names and shapes, but explicitly does not claim independently verified full-checkpoint generation in that environment. Three steps are an initial experiment, not a universal sampler recipe. Scheduler, guidance, adapter strength, and first/last-frame nodes still matter. A graph that accepts one reference picture does not automatically implement both endpoint frames.
For a useful local test, hold the prompt, seed, dimensions, frame count, sampler, and image inputs fixed while changing only the adapter. Watch motion and listen to speech. Record loading, encoding, sampling, decoding, and saving separately; a faster sampling stage can coexist with a slow complete job. These are proposed checks, not project measurements.
Why streaming is a separate implementation
The official FL2VA path lives under src/taomate-fl2av. It combines blockwise denoising, persistent clean audio-video KV state, endpoint guidance, and a specific runtime policy. The LoRA supplies learned deltas; the runtime manages continuity and scheduling. An ordinary three-step graph supplies neither automatically.
The documented FL2VA example uses eight GPUs with TP2/SP4, 24 fps, a 768×1376 canvas, recent-seven clean AV-KV, and whole-clip decoding. This is materially different from running a small adapter on a single desktop GPU. The broader T2AV requirements list Hopper/SM90 hardware, Linux, CUDA 12.8, PyTorch 2.8, and a validated eight-H20 configuration. Do not use a generic four-GPU runtime option as proof that every FL2VA example supports that configuration.
The duration labels are approximate
The official example maps its --time options to native frame counts:
| Option | Native frames | Actual duration at 24 fps |
|---|---|---|
--time 5 | 124 | 5.166667 seconds |
--time 10 | 260 | 10.833333 seconds |
--time 30 | 770 | 32.083333 seconds |
These are the documented example's geometries, not evidence that arbitrary requested lengths produce exact timelines. If your final delivery must be exactly thirty seconds, inspect the encoded output and decide how to trim picture and sound together. Avoid treating the nominal CLI option as the measured MP4 duration.
The card provides a checkpoint download and an infer.py invocation for its seated-presentation example. That command is an example entry point, not a documented general interface for every reader's two pictures. Inspect the subproject's accepted inputs before replacing assets or advertising a general FL2VA service.
Interpret the speed claims correctly
The model card also includes earlier T2AV benchmark numbers. They use a particular eight-H20 node and distinguish DiT time, final-chunk latent latency, and playable-video publication. Those figures are not a new FL2VA benchmark on an RTX card. Streaming latent availability also differs from a playable stream when a path performs whole-clip decoding.
For access today, choose the conversion when you want to investigate local LoRA loading; choose the official subproject when you can reproduce its distributed runtime. For a conventional online reference task, see MiniMax H3 Ref2VA, which has a different execution identity. For local file preparation, use the ComfyUI guide. Keep those routes separate from claims that TaoMate streaming has been integrated into H3Vid.