Uploaded November 2026 | Updated November 2026, 11 hours ago
MiniMax H3 in ComfyUI gets a new 6-step Pika H3 Flex workflow and Open-H3-Regenerate-2K refinement. I tested both on three scenes to see when 2K improves the video and when it changes things you never asked for.
I am using MiniMax H3 Singularity Ref2VA with the Pika H3 Flex Ref2VA BF16 LoRA. The workflow uses ManualSigmas with the 6-step schedule and LCM sampler. I also connected the MiniMax H3 latent upscaler with Open-H3-Regenerate-2K for an additional 3-step refinement pass.
But does a sharper video always mean a better result?
I tested three different scenes.
**Train fight:** I generated a 10-second action scene on a moving train. The 2K version looks sharper, but it changes the action. In the original, both characters jump between the train carriages. After the refinement, the man's jump disappears, and the action becomes slower.
**Drama scene:** Three women are having a conversation, with a red envelope reveal. The scene works well, but the 2K refinement adds unwanted gold to the plain envelope. Reducing the strength to 0.7 did not fix it. Changing the second sampler's start sigma to 0.8 fixed this problem in my test.
**Cup scene:** A woman picks up a cup and puts it back on the plate. This is where the 2K refinement gives a better-looking result. The shirt, face and wooden table have more detail. However, the refinement also changes her expression and adds a smile that I never requested.
My conclusion from these tests: 2K refinement is worth trying for slower scenes where extra texture and facial detail matter. For fast action scenes, I recommend staying with the original latent-upscaler workflow, based on the results I got.
I have not tested the 0.8 start sigma on the train fight yet, so I cannot say whether it fixes that action problem.
**WORKFLOW & DOWNLOADS**
My updated ComfyUI workflow:
patreon.com/ComfyuiWorkflowBlog/posts/minimax-h3-pika-174381818
Pika H3 Flex LoRAs (Ref2VA and FL2VA):
huggingface.co/Pika-Labs/Pika-H3-Flex
Open-H3-Regenerate-2K project:
nvlabs.github.io/Sana/Open-H3-Regenerate-2K
**IMPORTANT SETTINGS**
Pika H3 Flex Ref2VA:
- MiniMax H3 Singularity Ref2VA model
- Pika H3 Flex Ref2VA Comfy BF16 LoRA
- 6-step ManualSigmas schedule
- LCM sampler
- Do not stack an additional Turbo LoRA on this workflow
Open-H3-Regenerate-2K:
- Additional 3-step refinement
- Default start sigma: approximately 0.91
- Start sigma used for the envelope fix: 0.8
The settings and results shown here come from my tests. Different scenes may give different results.
If you try a start sigma of 0.8 on a fast action scene, let me know whether it keeps the motion and jumps correct. I would like to see how it works in other tests.
#MiniMaxH3 #ComfyUI #PikaH3Flex
MiniMax H3 in ComfyUI gets a new 6-step Pika H3 Flex workflow and Open-H3-Regenerate-2K refinement. I tested both on three scenes to see when 2K improves the video and when it changes things you never asked for.
I am using MiniMax H3 Singularity Ref2VA with the Pika H3 Flex Ref2VA BF16 LoRA. The workflow uses ManualSigmas with the 6-step schedule and LCM sampler. I also connected the MiniMax H3 latent upscaler with Open-H3-Regenerate-2K for an additional 3-step refinement pass.
But does a sharper video always mean a better result?
I tested three different scenes.
**Train fight:** I generated a 10-second action scene on a moving train. The 2K version looks sharper, but it changes the action. In the original, both characters jump between the train carriages. After the refinement, the man's jump disappears, and the action becomes slower.
**Drama scene:** Three women are having a conversation, with a red envelope reveal. The scene works well, but the 2K refinement adds unwanted gold to the plain envelope. Reducing the strength to 0.7 did not fix it. Changing the second sampler's start sigma to 0.8 fixed this problem in my test.
**Cup scene:** A woman picks up a cup and puts it back on the plate. This is where the 2K refinement gives a better-looking result. The shirt, face and wooden table have more detail. However, the refinement also changes her expression and adds a smile that I never requested.
My conclusion from these tests: 2K refinement is worth trying for slower scenes where extra texture and facial detail matter. For fast action scenes, I recommend staying with the original latent-upscaler workflow, based on the results I got.
I have not tested the 0.8 start sigma on the train fight yet, so I cannot say whether it fixes that action problem.
**WORKFLOW & DOWNLOADS**
My updated ComfyUI workflow:
patreon.com/ComfyuiWorkflowBlog/posts/minimax-h3-pika-174381818
Pika H3 Flex LoRAs (Ref2VA and FL2VA):
huggingface.co/Pika-Labs/Pika-H3-Flex
Open-H3-Regenerate-2K project:
nvlabs.github.io/Sana/Open-H3-Regenerate-2K
**IMPORTANT SETTINGS**
Pika H3 Flex Ref2VA:
- MiniMax H3 Singularity Ref2VA model
- Pika H3 Flex Ref2VA Comfy BF16 LoRA
- 6-step ManualSigmas schedule
- LCM sampler
- Do not stack an additional Turbo LoRA on this workflow
Open-H3-Regenerate-2K:
- Additional 3-step refinement
- Default start sigma: approximately 0.91
- Start sigma used for the envelope fix: 0.8
The settings and results shown here come from my tests. Different scenes may give different results.
If you try a start sigma of 0.8 on a fast action scene, let me know whether it keeps the motion and jumps correct. I would like to see how it works in other tests.
#MiniMaxH3 #ComfyUI #PikaH3Flex


![MiniMax H3 Low VRAM Upscaling in ComfyUI: 48 FPS, HDR & Best Settings(ComfyUI NVIDIA VFX Nodes)
I tested the new NVIDIA VFX nodes in ComfyUI to upscale my MiniMax H3 video on my [GPU name] with [X] GB of VRAM. Here are the best settings I found, plus how to get 48 fps and HDR.
The NVIDIA upscale took about 10 seconds. SeedVR2 took 3 to 4 minutes on the same video.
BEST SETTINGS (from my tests)
1. Generate the video with MiniMax H3
2. Latent Upscaler 3D (I used 1344 × 768). Do not skip this step.
3. RTX Video Super Resolution: STREAMING_ULTRA, scale 2
4. A second RTX Video Super Resolution node: DEBLUR_ULTRA
5. RTX Video Frame Generation: multiplier 2
6. Create Video: frame rate 48
7. Optional: RTX TrueHDR with HDR (HLG) or HDR PQ. Both looked good to me.
WHAT I FOUND
- The latent upscaler and the NVIDIA node do different jobs. In my test, the NVIDIA node did not replace the latent upscaler.
- ULTRA (the default) and BICUBIC looked almost the same.
- STREAMING_ULTRA gave the best result of the NVIDIA modes, even though NVIDIA made it for live streams and webcam video.
- Each node can use only one quality mode. To upscale and sharpen, you need two nodes.
- DENOISE and DEBLUR clean the video but do not make it bigger. On my video, DEBLUR_ULTRA helped and DENOISE_ULTRA showed no change.
- SeedVR2 still looks more natural, especially in the eyes, because it keeps a little noise. The NVIDIA nodes are much faster.
TWO MISTAKES TO AVOID
- 48 fps: the multiplier only adds frames. You also need to set Create Video to 48. If you leave it at 24, the video becomes twice as long and the audio ends halfway.
- HDR: ComfyUI does not save HDR information inside the video file. Export it as HDR (HLG or PQ) in your video editor before you upload. People with a normal screen will see a normal version.
LOW VRAM TIP
If you get an out-of-memory error, lower the scale, use a lower quality mode, or send fewer frames at once. Speed also depends on your graphics card.
REQUIREMENTS
- NVIDIA RTX graphics card (Frame Generation needs an RTX 40 series card or newer)
- PyTorch for CUDA 13 or newer
- Install from ComfyUI Manager: search for ComfyUI Nvidia VFX Nodes (the old name was Nvidia RTX Nodes)
Workflow: https://www.patreon.com/ComfyuiWorkflowBlog/posts/minimax-h3-for-172802597
ComfyUI Nvidia VFX Nodes: https://github.com/Comfy-Org/ComfyUI_Nvidia_VFX_Nodes
Which one do you use for your AI videos, SeedVR2 or the NVIDIA nodes? Tell me in the comments.
#MiniMaxH3 #ComfyUI #AIVideo MiniMax H3 Low VRAM Upscaling in ComfyUI: 48 FPS, HDR & Best Settings(ComfyUI NVIDIA VFX Nodes)](https://i.ytimg.com/vi/DBGBD6GaPQA/mqdefault.jpg)







