Uploaded September 2026 | Updated November 2026, 1 day ago
Alibaba PAI released an 8-step acceleration LoRA for MiniMax H3. I tested it in
ComfyUI against the 4-step turbo LoRA I have been using, on the same scenes and the
same prompts.
In the video:
Fast motion and a fighting scene on both LoRAs, same prompt
What happens to sound on 4 steps
A prompt the 4-step result ignores
One type of scene where this model still fails
Generation times on a 5090, and how I cut mine down using both LoRAs in one workflow
Timings, settings and the full node setup are shown on screen.
WHERE THE FILE GOES
This one does not go into your loras folder. It goes into ComfyUI/models/pdd_acc/.
If that folder is missing, create it with that exact name.
Two files were released, one for FL2VA and one for Ref2VA. Match the file to the
model you are running. The wrong file gives a wrong result and it will look like the
LoRA is broken.
LINKS
Custom node: github.com/Jalen-Brunson/ComfyUI-MiniMax-H3-PDD-Acc
Official LoRAs from Alibaba PAI: huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/tree/main
4-Step Turbo LoRA source: huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main
Workflow: patreon.com/ComfyuiWorkflowBlog/posts/minimax-h3-8-my-168587999
#MiniMaxH3 #ComfyUI #AIVideo
Alibaba PAI released an 8-step acceleration LoRA for MiniMax H3. I tested it in
ComfyUI against the 4-step turbo LoRA I have been using, on the same scenes and the
same prompts.
In the video:
Fast motion and a fighting scene on both LoRAs, same prompt
What happens to sound on 4 steps
A prompt the 4-step result ignores
One type of scene where this model still fails
Generation times on a 5090, and how I cut mine down using both LoRAs in one workflow
Timings, settings and the full node setup are shown on screen.
WHERE THE FILE GOES
This one does not go into your loras folder. It goes into ComfyUI/models/pdd_acc/.
If that folder is missing, create it with that exact name.
Two files were released, one for FL2VA and one for Ref2VA. Match the file to the
model you are running. The wrong file gives a wrong result and it will look like the
LoRA is broken.
LINKS
Custom node: github.com/Jalen-Brunson/ComfyUI-MiniMax-H3-PDD-Acc
Official LoRAs from Alibaba PAI: huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/tree/main
4-Step Turbo LoRA source: huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main
Workflow: patreon.com/ComfyuiWorkflowBlog/posts/minimax-h3-8-my-168587999
#MiniMaxH3 #ComfyUI #AIVideo


![WAN 2.2 Animate in ComfyUI: InfiniteTalk + Uni3C — Low VRAM (GGUF Q4) [Free Workflow]
Download the free workflow: https://aistudynow.com/wan-2-2-animate-comfyui-infinitytalk-uni3c-q4-low-vram/
#ComfyUI #WAN22 #InfinityTalk
Swap any face or object inside one ComfyUI flow. Blend looks real. Lips match. Camera move support with Uni3C. FP8 vs GGUF Q4 test on low VRAM.
what you will learn
InfiniteTalk with WAN 2.2 Animate for better lip sync
Uni3C camera motion test and when to keep it off
FP8 vs Q4 GGUF on low VRAM and what to pick
LoRA we used
Relighting LoRA for Animate to match scene color and tone WanAnimate_relight_lora_fp16.safetensors
LightX 4-step LoRA set for Animate or I2V
low_noise_model.safetensors
Pusa v1 motion LoRA if you want stronger motion : Wan21_PusaV1_LoRA_14B_rank512_bf16.safetensors
HPS LoRA for human-preference quality tuning : Wan2.2-Fun-A14B-InP-LOW-HPS2.1_resized_dynamic_avg_rank_15_bf16.safetensors
upload the edit video and pick a supported WAN 2.2 size with Resolution Master
upload your subject image at the same size, remove background
mask with points: green keep, red remove, test mask once
models:
Wan2.2-Animate-14B FP8 model: Wan2_2-Animate-14B_fp8_e4m3fn_scaled_KJ.safetensors
Folder: ComfyUI/models/diffusion_models
Optional for low VRAM: Wan2.2 I2V GGUF Q4 or Q5 pairs here is the link
disclosure and credits
this video uses AI to alter real footage for an educational demo. altered or synthetic content label is on in Studio.
WAN, InfinityTalk, Uni3C, LoRA models belong to their creators. credits to them. WAN 2.2 Animate in ComfyUI: InfiniteTalk + Uni3C — Low VRAM (GGUF Q4) [Free Workflow]](https://i.ytimg.com/vi/fB7rCasr6yQ/mqdefault.jpg)







