Uploaded September 2026 | Updated November 2026, 1 day ago
I tested Singularity, a community fine-tune of MiniMax H3, against the official Ref2VA model in ComfyUI with the same seed, and I'm sharing my updated 8-step workflow. The result: sharper texture, better faces, and the strongest fight scenes I've made so far.
Singularity is not an official MiniMax model. It comes in two INT8 versions: 34 GB and 21 GB. If you have low VRAM, use the 21 GB one. I ran everything on an RTX 5090 with 32 GB of VRAM.
In this video:
- Official MiniMax H3 Ref2VA INT8 vs Singularity, same seed
- My 8-step workflow with two LoRAs
- How I split my character sheet into separate face, hand, and sword references
- A quick 4-step workflow, and a 4-step + second sampler + upscale version that looks almost as good as 8 steps
- A 15-second demon hunter vs. giant oni fight in a Kyoto street
- An armored flying warrior vs. a dragon in a thunderstorm
Resources
Workflow (optimized): patreon.com/ComfyuiWorkflowBlog/posts/singularity-h3-3-169980649
Singularity model (Hugging Face): huggingface.co/WarmBloodAban/Minimax-h3_Singularity
If you want more MiniMax H3 tests like this, subscribe.
#MiniMaxH3 #ComfyUI #AIVideo
I tested Singularity, a community fine-tune of MiniMax H3, against the official Ref2VA model in ComfyUI with the same seed, and I'm sharing my updated 8-step workflow. The result: sharper texture, better faces, and the strongest fight scenes I've made so far.
Singularity is not an official MiniMax model. It comes in two INT8 versions: 34 GB and 21 GB. If you have low VRAM, use the 21 GB one. I ran everything on an RTX 5090 with 32 GB of VRAM.
In this video:
- Official MiniMax H3 Ref2VA INT8 vs Singularity, same seed
- My 8-step workflow with two LoRAs
- How I split my character sheet into separate face, hand, and sword references
- A quick 4-step workflow, and a 4-step + second sampler + upscale version that looks almost as good as 8 steps
- A 15-second demon hunter vs. giant oni fight in a Kyoto street
- An armored flying warrior vs. a dragon in a thunderstorm
Resources
Workflow (optimized): patreon.com/ComfyuiWorkflowBlog/posts/singularity-h3-3-169980649
Singularity model (Hugging Face): huggingface.co/WarmBloodAban/Minimax-h3_Singularity
If you want more MiniMax H3 tests like this, subscribe.
#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)





