MiniMax H3 native workflows: fl2va and ref2va serve different tasks

File and download boundary: ComfyUI World provides source references and metadata only. Model and LoRA weights, node packages, and workflow files stay with their original providers; this site does not store, mirror, or proxy them. Downloads and access conditions are controlled by each provider.

Purpose and scope

This reference covers the open-weight native ComfyUI route, not remote Partner/API execution. Local operation depends on component pairing, core version and actual hardware.

Source-supported context

The official overview specifies ComfyUI 0.30.0 or later. The current native-workflow guide distinguishes fl2va for T2V/I2V from ref2va for R2V, lists a shared text encoder and separate video/audio VAEs, and documents task-specific turbo LoRAs. The August 3, 2026 support announcement is historical, not a local benchmark. Overview Native workflows

Before you start

1. Text or first/last-frame image-to-video uses the corresponding fl2va route.

2. Multimedia-reference generation uses ref2va and its reference-input rules.

3. Local execution, cloud-hosted ComfyUI and remote API services differ in location, authorization and cost.

Steps and checks

The sequence below is an editorial procedure based on the cited context, not a diagnosis already confirmed for your environment.

Step 1. Select a task template from the native guide and check core/node registration. Updating only the frontend does not add backend model support.

Step 2. Select the task's diffusion file rather than treating fl2va and ref2va as two mandatory sequential stages. Supply the text encoder and current video/audio VAEs. Check the selected template's LoRA references: turbo mode is optional, but some template model scans still expect the matching LoRA file.

Step 3. Start with the selected template's supported preview settings to validate the complete audiovisual path, including resolution selection and audio connections to saving. Change dimensions or duration only within that template's supported controls.

Step 4. Add optional acceleration such as SageAttention only after the baseline passes. Route attention imports and silent output to their respective diagnostics.

Completion check

The task uses the correct diffusion route and inputs, and both video and audio output are inspectable.

Limitations and cautions

The two diffusion files serve different routes and are not both loaded for every task. The turbo LoRAs are task-specific; their presence in a template does not make turbo mode mandatory. No 8GB, 24GB or other hardware claim is verified here. Open weights are not blanket use rights: check the selected repository's terms, location and commercial-use requirements against the upstream license and ComfyUI overview.

Files named by the source documentation

These are current native-template declarations across T2V/I2V and R2V, not inspected local files or direct downloads. The final two entries are route-specific turbo LoRAs; check whether the selected template scans for them even when turbo is disabled. No weights or checksum results are included; verify the revision for your task.

File	Directory category	Documented source
minimax_h3_fl2va_pruned_int8_convrot.safetensors	diffusion_models	Source
minimax_h3_ref2va_pruned_int8_convrot.safetensors	diffusion_models	Source
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors	text_encoders	Source
minimax_h3_video_vae_int8_convrot.safetensors	vae	Source
minimax_h3_audio_vae_fp32.safetensors	vae	Source
minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors	loras	Source
minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors	loras	Source

Original sources

• MiniMax H3 overview — original record checked 2026-09-21.
• MiniMax H3 native workflows — rechecked 2026-09-25.
• MiniMax H3 Day-0 support in ComfyUI — original record checked 2026-09-21.

Localization prepared: 2026-09-22. The source-check dates above were inherited from the existing article; this translation does not claim they were all rechecked today. No GPU run, minimum-memory measurement or runtime guarantee is supplied.

Related troubleshooting and guides

• SageAttention or Triton import failure: match the platform and build combination
• VHS output is nearly silent: inspect audio dtype before increasing gain
• Video export fails or plays incorrectly: check frames, encoding and audio separately