Modelpublic

fbjr/MiniMax-H3-Acc-LoRAs-sidecar

sourceHugging Faceotherupdated 21d agoView on Hugging Face

6likes8.3kdownloads

Model Card

MiniMax-H3 acceleration LoRAs for ComfyUI

MiniMax H3 normally takes 16 sampler steps to render a clip with sound. With these files and one ComfyUI node it takes 5 to 8, at close to the same quality. They are conversions of alibaba-pai/MiniMax-H3-Acc-LoRAs, the official Parallel Decoding Distillation LoRAs for MiniMaxAI/MiniMax-H3. Nothing was trained here.

Each file is a sidecar: it rides alongside your H3 checkpoint and carries three things the acceleration needs. A LoRA update for the transformer, an update to the timestep modulation, and a bank of 32 replacement output heads, one per slice of the trajectory, which is the actual trick. A stock LoRA loader applies only the first and silently renders at the wrong quality. The MiniMaxH3PDDLoRA node here applies all three and gives your sampler the schedule the heads were made for. Take the node and the files from the same revision; they are checked against each other.

What changed since the last update

2026-09-03	what it means for you
The four files were regenerated. Weights are identical; each now carries a fingerprint of the checkpoint it was made for.	The node refuses a file made for the wrong checkpoint instead of rendering something plausible at the wrong quality. Update the node folder and the files together; old and new do not mix.
New optional input `unmerged_blocks`, off by default.	An accuracy knob for quantised checkpoints, at a render-time cost. Ignore it unless you are experimenting.

Quick start

1.Read your checkpoint's filename. It says fl2va or ref2va, and it says pruned if it is. That is enough to pick a file below.
2.Put that file in models/loras/.
3.Put the folder comfyui_minimax_h3_pdd/ in custom_nodes/ and restart. Nothing beyond core ComfyUI; you do not need the source repo. If you already run ComfyUI-h3-explorations, you have this node; do not install it twice. If it fails to load with No module named 'pdd_observe', download it again: the first 2026-09-03 copy was incomplete.

   hf download fbjr/MiniMax-H3-Acc-LoRAs-sidecar --include "comfyui_minimax_h3_pdd/*" --local-dir ComfyUI/custom_nodes

1.Open a graph from workflows/, or wire the node yourself:

UNETLoader ──► MiniMaxH3PDDLoRA ──MODEL──► BasicGuider ──► SamplerCustomAdvanced
                      └──────────SIGMAS───────────────────────────┘

1.Set steps on the node. 5 is a good default; 8 is the distilled count.

Two rules: no BasicScheduler, wire the node's SIGMAS to the sampler instead, since a scheduler at the same count gives different sigmas than the heads were fused for; and set steps only on the node, which refuses a count it cannot fuse.

Which file

file	for	inside
`minimax_h3_fl2va_pdd_8step_comfy.safetensors`	pruned FL2VA	backbone LoRA, head bank, modulation update pre-solved into the pruned checkpoint's rank-8 basis
`minimax_h3_fl2va_pdd_8step_adaln2688_comfy.safetensors`	any FL2VA	the same, with the modulation update at full width
`minimax_h3_ref2va_pdd_8step_comfy.safetensors`	pruned Ref2VA	as the pruned FL2VA one, from the Ref2VA LoRA
`minimax_h3_ref2va_pdd_8step_adaln2688_comfy.safetensors`	any Ref2VA	as the FL2VA `adaln2688` one, from the Ref2VA LoRA

Not sure whether yours is pruned? Take adaln2688; it works on both. The two partitions share identical key sets, which is why each file carries a checkpoint fingerprint: without it a wrong pairing loads with no error.

Steps

The heads were distilled over a 32-point grid in blocks of width 4; a count is legal when its blocks tile the grid at widths 4 to 8.

steps	blocks	versus 8
4	8, 8, 8, 8	noticeably coarser; the last step covers most of the trajectory, and no other 4-block layout is legal
5	8, 8, 8, 4, 4	within 2%
6	8, 8, 4, 4, 4, 4	within 1%
7	8, 4, 4, 4, 4, 4, 4	within 1%
8	4 x 8	the distilled count

5 to 8 are a plateau; pick by time. 1, 2, 16 and 32 are accepted but untested; everything else is refused. Different counts give different clips, not better and worse versions of one.

Workflows

Four text-to-video-with-audio graphs on FL2VA, one dialogue scene. Repoint the loaders at your checkpoint and VAEs.

graph	why
`t2va_pdd_5step.json`	start here
`t2va_pdd_8step.json`	the distilled count
`t2va_pdd_4step.json`	fastest and visibly coarser, so you can see why
`t2va_pdd_8step_heads_off.json`	control: the checkpoint's own output heads

Knobs

Leave everything but steps alone unless you are experimenting.

input	default	what it does
`steps`	8	the step count, and what `SIGMAS` emits
`strength`	1.0	scales the whole file at once
`patch_heads`	on	the replacement output heads; off is a control
`head_strength`	-1, follow `strength`	scales the output heads alone; 0 runs the checkpoint's own
`nfe`	0	experiment only; 0 is the ordinary mode
`unmerged_blocks`	empty	accuracy knob for quantised checkpoints: named blocks apply their LoRA update at the call instead of merging it into the quantised weight
`unmerged_strength`, `unmerged_window`	follow, empty	experiment only; inert unless `unmerged_blocks` is set

Versions

Converter 3, regenerated 2026-09-03. The node folder's own README.md names the commit it was built from. Content hashes over sorted metadata and every tensor's bytes in key order:

file	sha256 prefix
`minimax_h3_fl2va_pdd_8step_comfy.safetensors`	`115fcd82dbf2125a`
`minimax_h3_fl2va_pdd_8step_adaln2688_comfy.safetensors`	`cd788e5e3ac7de3f`
`minimax_h3_ref2va_pdd_8step_comfy.safetensors`	`102a83d1f57c1c2f`
`minimax_h3_ref2va_pdd_8step_adaln2688_comfy.safetensors`	`7c402b48fc2a6b85`

Under the hood

The published files use diffusers-side key names, so ComfyUI skips every tensor and renders an ordinary undistilled pass that looks like a weak LoRA. The conversion fuses q, k and v into ComfyUI's single projection with alpha scaled to match, renames two modules and swaps the SwiGLU halves, ships the 32 output heads verbatim and fuses them per block at load so one file serves every legal count, and carries the modulation update in full in the adaln2688 files or pre-solved into the pruned checkpoint's rank-8 basis in the others. If you write your own loader: the heads are verbatim, not deltas from head 0, which is the opposite of ComfyUI core's convention.

Limitations

Four steps cannot be improved, since 8, 8, 8, 8 is the only legal layout. Block width 8 is twice the trained width and unmeasured. Coarser steps degrade video and audio together; you hear it first. This is an 8-step distillation, and far from 8 is extrapolation.

Credits

—MiniMax for MiniMax H3.
—alibaba-pai for the Acc LoRAs these are converted from, and for the PDD method.
—The node is generated from `ComfyUI-h3-explorations`, which is where to file issues and where the rest of the H3 work lives.
—silveroxides, whose ComfyUI-UtilsCollection PDD implementation separates LoRA strength from head strength; the node that reads these files borrows that split.

License

These are Model Derivatives of MiniMax H3 and are distributed under the MiniMax H3 Community License Agreement, the same terms as both upstream releases. The full agreement is in `LICENSE`; the required attribution is in `NOTICE`.

Two terms worth reading before you use these commercially or redistribute them, neither of which is unusual but both of which are easy to miss:

—The agreement is territorially limited — worldwide excluding the Excluded Territories it names.
—Commercial products or services generating over 20 million USD in yearly revenue need separate written authorisation from MiniMax.

If you redistribute these files or anything derived from them, the agreement requires you to pass on a copy of it, carry prominent notices on modified files, and include the NOTICE text. You may not relicense them under different terms.

MiniMax H3 is licensed under the MiniMax H3 Community License Agreement,
Copyright (c) 2026 MiniMax. All Rights Reserved.