VDN-H3 Turbo Goes Standalone: 8-Step LoRA for MiniMax H3

ComfyUI Wikinews

drbaph extracted the VDN-H3 Turbo adapter into a standalone 8-step LoRA for MiniMax H3, with pruned fl2va/ref2va variants plus an INT8 ConvRot VDN checkpoint for Comfy Kitchen.

VDN-H3 is the hybrid-attention Video DeltaNet rework of MiniMax H3 that replaces quadratic long-range attention with a linear Video Delta Attention branch (read the original coverage). On September 9, community contributor drbaph published two follow-ups on Hugging Face: the VDN-H3 Turbo adapter extracted into a standalone 8-step LoRA usable on unmodified MiniMax H3 checkpoints, and a pre-quantized INT8 ConvRot version of the VDN-H3 checkpoint that cuts its VRAM footprint by roughly 1.8 GB and speeds up the linear branch about 2.7x.

Why standalone matters

Until now the only way to run the VDN-H3 Turbo adapter was through the ComfyUI-VDN-H3 custom nodes with the full hybrid-attention patch: the 8-step checkpoint, its linear branch, and both adapters loaded together. The new standalone LoRA strips the Turbo adapter out of that stack and repackages it as a plain MiniMax H3 LoRA, so it loads with ComfyUI's built-in LoRA loader on the stock H3 checkpoint, no custom node required.

Per drbaph, the Turbo adapter inside VDN-H3 was initialized from larryvrh's MiniMax H3 Turbo LoRA and then further DMD-trained as part of VDN-H3. This release is the standalone conversion of that resulting adapter, so it carries distillation signal from both lineages.

What was released

The LoRAs live in the experimental/ folder of drbaph/MiniMax-H3-Turbo-Lora-ComfyUI:

FileBaseNotes
minimax_h3_dmd_8step_turbo.safetensorsnon-pruned fl2va / ref2vaone LoRA for both unpruned variants
minimax_h3_dmd_fl2va_8step_turbo_pruned.safetensorspruned fl2vaseparate variant for the pruned/curve-form checkpoint
minimax_h3_dmd_ref2va_8step_turbo_pruned.safetensorspruned ref2vaseparate variant for the pruned/curve-form checkpoint

Recommended settings: 8 steps, strength 0.65 to 1.0. Each pruned file is about 2.1 GB.

The INT8 ConvRot VDN checkpoint

The second release targets the full VDN-H3 route. drbaph/vdn-minimax-h3-int8-convrot-comfyui pre-quantizes the 8-step stage-dmd-step-250 checkpoint into Comfy Kitchen INT8 ConvRot format, the same serialization ComfyUI ships for its own int8-convrot diffusion models. Only the linear branch's dispatched matmul weights are quantized; adapters, biases, norms and convolutions pass through untouched.

bf16 originalINT8 ConvRot
linear branch4.28 GB2.20 GB
stage total5.09 GB3.25 GB
branch matmuls1x~2.7x faster
end-to-end (single test run)~111 s~95 s

An A/B render at 1280x736 / 61 frames with the same seed was visually identical, and peak VRAM while loading dropped by roughly 4.7 GB. The quantized stage requires ComfyUI-VDN-H3 v1.3.0 or newer and drops into ComfyUI/models/vdn/ like the original.

Ref2VA example rendered with the INT8 ConvRot stage: 8 steps, er_sde / beta, 928x928.

How it fits with the other H3 accelerators

Community testing in the Banodoco H3 channels this week put the numbers in context: the VDN route trades some speed against the lightx2v turbo LoRAs (one benchmark timed the 8-step VDN Turbo at 2:04 for 1280x736 versus 1:24 for the 4-step lightx2v turbo on the same hardware) but holds fast motion better than the FastH3 route. The standalone 8-step LoRA now gives a middle option: VDN's DMD-distilled adapter without loading the linear branch at all.

For 50-step audio work, drbaph's suggested recipe is to generate at 512x512 minimum and upscale to 1080p through the ref2va turbo 8-step path, downscaling to 256 first for speed.

Availability

Comments

Sign in with GitHub to join the discussion.

Loading comments…
VDN-H3 Turbo Goes Standalone: 8-Step LoRA for MiniMax H3 | ComfyUI Wiki