中

Home / LoRA library / Alibaba PAI PDD Acc 8-step (official distillation)

Alibaba PAI PDD Acc 8-step (official distillation): Compatible GPUs, Download & Arena Rank

Step-distilled · 8 steps · updated 2026-09-28

Blind arena: i2v #14 · Elo 968.7 (3950 votes) · t2v #9 (source: H3 Acceleration Arena)
ItemValue
Steps8 steps · ~2.1× faster
Quality costLow (official distillation)
RequiresComfyUI>=0.33, ComfyUI-MiniMax-H3-PDD-Acc nodes, dedicated sigmas
Conflictsnone
Includes a Parallel Decoding Distillation head bank; a regular LoRA loader silently drops the distillation head; CFG=1.0, euler + sigmas output by the node

Representative GPUs: 864 × 480 · 8s

Combo: pruned INT8 + NVFP4 encoder + this acceleration + SageAttention, 32GB RAM / fast-disk.

GPUVRAM864×480 · 8sNote
RTX 4060 8GB8GB⚠ 7 min 12 sVRAM peak 7G
RTX 3060 12GB12GB6 min 50 sVRAM peak 11G
RTX 4070 12GB12GB3 min 59 sVRAM peak 11G
RTX 4070 Ti Super 16GB16GB2 min 54 sVRAM peak 15G
RTX 4080 Super 16GB16GB2 min 36 sVRAM peak 15G
RTX 4090 24GB24GB1 min 53 sVRAM peak 23G
RTX 5090 32GB32GB1 min 22 sVRAM peak 26G
RTX PRO 6000 Blackwell 96GB96GB1 min 17 sVRAM peak 47G

Which GPUs run 1344×768 · 5s

GTX 1060 6GB GTX 1070 8GB GTX 1080 8GB GTX 1080 Ti 11GB RTX 2060 6GB RTX 2060 12GB RTX 2060 Super 8GB RTX 2070 Super 8GB RTX 2080 Ti 11GB RTX 2080 Ti 22GB (modded) RTX 3050 8GB RTX 3060 8GB RTX 3060 12GB RTX 3060 Ti 8GB RTX 3070 8GB RTX 3070 16GB (modded) RTX 3070 Ti 8GB RTX 3080 10GB +30 more

→ Check your own GPU in the calculator (this option pre-selected)

FAQ

How much VRAM does Alibaba PAI PDD Acc 8-step (official distillation) need?

By our estimates, 1344×768 · 5s needs roughly 6GB+ of VRAM (48 of 49 listed GPUs workable). Small clips at 736×416·5s run on even less.

Can it be combined with other acceleration?

Pick one step-distilled option only — they change the sampling trajectory. Attention backends (Sage/Kitchen/SLA/VSA) and caches stack on top. PDD requires its dedicated node.

Where do I download it?

See the source links below — we only link to the official author repositories and never rehost files.

Sources

Related: LoRA blind-test leaderboard · Acceleration compared · All acceleration options