Uploaded October 2026 | Updated November 2026, 4 days ago
I tested MiniMax H3 Singularity V2 in ComfyUI on fast action, a 27-second continuous camera shot, and a 30-second fight between two characters. You will see what got better, where it still struggles, and the exact LoRAs I used.
Singularity is a community fine-tune of MiniMax H3. According to its creator, V2 was trained on more action scenes and more complex movement. The creator also says it follows prompts more closely, fixes the audio reference problem from the older version, and gives better realism and CG.
What I found in my tests:
• Fast action is much better. When the characters jump and hit, you can feel the power. I could not get this from older versions.
• Camera movement improved a lot. With all LoRAs turned off, I made one continuous shot of about 27 seconds.
• The 30-second rooftop fight looks much better than the previous version, but it is not perfect. In some runs the same hit repeated, and the model did not always understand the dodge. It took me several prompt versions to get this result.
• I also show a V1.3 vs V2 comparison from Reddit. Some people still prefer V1.3 for calm character shots, but for action, V2 is much better in my tests.
Workflow
patreon.com/ComfyuiWorkflowBlog/posts/minimax-h3-v2-173657891
Model
MiniMax H3 Singularity V2, full model (about 34 GB), by WarmBloodAban:
huggingface.co/WarmBloodAban/Minimax-h3_Singularity
Smaller pruned V2 versions (21 GB and 12.5 GB), by Abiray:
huggingface.co/Abiray/Minimax-h3-Singularity-V2-Pruned
How the workflow works
Each part has its own group, so you can turn it on and off.
1. Audio refine (off by default). It makes the sound better but adds 6 more sampling steps. It uses ComfyUI-H3-AudioRefine:
github.com/Adudeguyman/ComfyUI-H3-AudioRefine
2. Turbo LoRA: LightX2V 8-step v1.0 at 768p. LightX2V keeps updating these LoRAs, so check their page for the newest version:
huggingface.co/lightx2v/Minimax-h3-Turbo
3. Scene LoRAs. Do not turn them all on at once. Pick the ones that match your scene. In the second pass, keep the strength lower, around 0.3 to 0.4.
4. Upscaling group at the end.
Scene LoRAs
• GunFu, for gunfights:
huggingface.co/JOKER141/MiniMaxH3-GunFu-Gunfight-LORA
• SMACK, for hits and smashes with stronger impact (no trigger word):
huggingface.co/LeechTM/SMACK
• Motion Repair, fixes messy or broken motion. Its creator says it works best together with a combat LoRA:
huggingface.co/JOKER141/MiniMax-H3-General-Motion-Continuity-Repair
• Bunny Weapon Combat, for weapon fights (trigger word: BUNNY):
huggingface.co/JOKER141/MiniMax-H3-Weapon-Combat-LoRA
• H3 Fight Flow V1, for fights between two characters. This is the only LoRA I used in the rooftop fight. Trigger words: Unarmed combat interaction, Melee weapon interaction, or Gun-fu interaction, depending on the fight:
huggingface.co/JOKER141/H3-Fight-Flow-Fight-Interaction-LoRA
• Facial Realism Close-Up, for close-up faces (trigger word: Facial Realism):
huggingface.co/prithivMLmods/MiniMax-H3-Facial-Realism-CloseUp
• H3 Realism People, for realistic people (trigger word: r34l1sm):
huggingface.co/fal/MiniMax-H3-Realism-People-LoRA
If you try the workflow, tell me in the comments which GPU you used and which LoRA combination worked best for your scene.
All video clips in this video were generated with AI.
#MiniMaxH3 #ComfyUI #AIVideo
I tested MiniMax H3 Singularity V2 in ComfyUI on fast action, a 27-second continuous camera shot, and a 30-second fight between two characters. You will see what got better, where it still struggles, and the exact LoRAs I used.
Singularity is a community fine-tune of MiniMax H3. According to its creator, V2 was trained on more action scenes and more complex movement. The creator also says it follows prompts more closely, fixes the audio reference problem from the older version, and gives better realism and CG.
What I found in my tests:
• Fast action is much better. When the characters jump and hit, you can feel the power. I could not get this from older versions.
• Camera movement improved a lot. With all LoRAs turned off, I made one continuous shot of about 27 seconds.
• The 30-second rooftop fight looks much better than the previous version, but it is not perfect. In some runs the same hit repeated, and the model did not always understand the dodge. It took me several prompt versions to get this result.
• I also show a V1.3 vs V2 comparison from Reddit. Some people still prefer V1.3 for calm character shots, but for action, V2 is much better in my tests.
Workflow
patreon.com/ComfyuiWorkflowBlog/posts/minimax-h3-v2-173657891
Model
MiniMax H3 Singularity V2, full model (about 34 GB), by WarmBloodAban:
huggingface.co/WarmBloodAban/Minimax-h3_Singularity
Smaller pruned V2 versions (21 GB and 12.5 GB), by Abiray:
huggingface.co/Abiray/Minimax-h3-Singularity-V2-Pruned
How the workflow works
Each part has its own group, so you can turn it on and off.
1. Audio refine (off by default). It makes the sound better but adds 6 more sampling steps. It uses ComfyUI-H3-AudioRefine:
github.com/Adudeguyman/ComfyUI-H3-AudioRefine
2. Turbo LoRA: LightX2V 8-step v1.0 at 768p. LightX2V keeps updating these LoRAs, so check their page for the newest version:
huggingface.co/lightx2v/Minimax-h3-Turbo
3. Scene LoRAs. Do not turn them all on at once. Pick the ones that match your scene. In the second pass, keep the strength lower, around 0.3 to 0.4.
4. Upscaling group at the end.
Scene LoRAs
• GunFu, for gunfights:
huggingface.co/JOKER141/MiniMaxH3-GunFu-Gunfight-LORA
• SMACK, for hits and smashes with stronger impact (no trigger word):
huggingface.co/LeechTM/SMACK
• Motion Repair, fixes messy or broken motion. Its creator says it works best together with a combat LoRA:
huggingface.co/JOKER141/MiniMax-H3-General-Motion-Continuity-Repair
• Bunny Weapon Combat, for weapon fights (trigger word: BUNNY):
huggingface.co/JOKER141/MiniMax-H3-Weapon-Combat-LoRA
• H3 Fight Flow V1, for fights between two characters. This is the only LoRA I used in the rooftop fight. Trigger words: Unarmed combat interaction, Melee weapon interaction, or Gun-fu interaction, depending on the fight:
huggingface.co/JOKER141/H3-Fight-Flow-Fight-Interaction-LoRA
• Facial Realism Close-Up, for close-up faces (trigger word: Facial Realism):
huggingface.co/prithivMLmods/MiniMax-H3-Facial-Realism-CloseUp
• H3 Realism People, for realistic people (trigger word: r34l1sm):
huggingface.co/fal/MiniMax-H3-Realism-People-LoRA
If you try the workflow, tell me in the comments which GPU you used and which LoRA combination worked best for your scene.
All video clips in this video were generated with AI.
#MiniMaxH3 #ComfyUI #AIVideo




![FLUX.2 Klein ComfyUI Workflow: Image Edit New Trick & ControlNet
If your FLUX.2 Klein results look like plastic or your GPU is struggling, you have to watch this. I created a new Flux Klein ComfyUI Workflow that fixes the realism issues and runs on low VRAM. I also found a secret ControlNet Hack, image editing trick that nobody know
In this video, I break down exactly how to use the new NVFP4 model format. It cuts VRAM usage by 55 percent, making it perfect for 4090 and 5090 cards. If you have an 8GB card, I show you how to use the GGUF model to get the same high quality.
You will also learn how to unlock the hidden Beta 57 scheduler by installing the RES4LYF node, which is the key to fixing that plastic AI look. Finally, I show you my dedicated image editing workflow to change clothes and transfer poses perfectly.
DFans – AI Influencer Monetization Platform
https://dfans.co/?ref=LA63HR
invitation code:- LA63HR
Files And Workflows You Need To Download:- https://aistudynow.com/flux-2-klein-fix-crashes-run-9b-on-6gb-vram-workflow-download/
Monetize Your ComfyUI Content [Link to your monetization offer]
#fluxklein #comfyui #comfyuiworkflow FLUX.2 Klein ComfyUI Workflow: Image Edit New Trick & ControlNet](https://i.ytimg.com/vi/jV7SNCMEKMw/mqdefault.jpg)





