FastH3 8-Step V2 Gets Native ComfyUI Templates

When FastH3 Preview v1 launched in late August, running it inside ComfyUI was not realistic: the VSA adapters needed FastVideo's own launchers, and we pointed ComfyUI users at distillation LoRAs like H3 Turbo-SLA and PDD Acc LoRA instead. That has changed. FastVideo shipped an 8-Step V2 checkpoint, Comfy-Org repacked it for ComfyUI, and the official workflow template repo now carries two native FastH3 templates.

The official FastH3 text-to-video template in ComfyUI's template browser.

The 8-Step V2 checkpoint

FastVideo-FastH3-8-Step-V2 is a data-free DMD2 distillation of MiniMax H3 trained with VSA-H3 attention at 80% sparsity. It generates synchronized video and audio in 8 transformer forwards instead of the base model's full schedule, and this generation ships with a permissive change: the checkpoint's own README documents the attention requirements directly, so it can be served outside the FastVideo stack as long as the sparse attention pattern is reproduced at inference.

Comfy-Org repacked the distilled transformer as fastvideo_fasth3_8step_v2_pruned_int8_convrot.safetensors in the Comfy-Org/FastVideo-FastH3 repo, which redirects to FastVideo's FastH3-Comfy; the base-model text encoder and the video/audio VAE files live on Comfy-Org/MiniMax-H3.

Two official templates

Template version 0.11.61 of Comfy-Org/workflow_templates added both workflows on September 15:

• video_fastvideo_fasth3_t2v.json: text-to-video-and-audio. The MiniMaxH3ImageToVideo node runs with no images connected.
• video_fastvideo_fasth3_i2v.json: text-to-video plus first/last-frame image conditioning. Connect first_frame and/or last_frame for fl2va.

The first/last-frame variant of the FastH3 template.

Both templates share one critical detail: the checkpoint was trained with VSA-H3 attention at 80% sparsity, and the templates wire in a BlockSparseAttention node set to the vsa method with keep_percent 10 to reproduce that pattern. The sol-attn and sla sparse methods are not compatible with this checkpoint. Resolution stays on H3's native 768px short-edge canvas, and duration snaps to the model's 17-frames-per-block grid at 24 FPS.

Download FastH3 t2v template Download FastH3 i2v template

Scope

The distilled checkpoint covers text-to-video-and-audio and first/last-frame conditioning only. Ref2VA (multi-reference conditioning) was not distilled, so reference-based tasks still need the base MiniMax H3 model. An added bonus of this generation: FastVideo's comfy-kitchen RoPE path means the templates run on standard ComfyUI nightly builds, not the FastVideo stack.

For context on where FastH3 sits in the wider acceleration landscape, see the H3 Acceleration Arena blind A/B ranking.