Uploaded June 2025 | Updated November 2026, 14 hours ago
In this video, you'll learn:
• how to install gemma 3n on android Mobile (even though it's not on the Play Store yet)
• Why this offline AI is different from cloud-dependent tools like ChatGPT or Gemini
• The surprising performance of the E4B and E2B models – including where they shine and where they fall short
I got early access to the developer preview, and let me just say – the results were a mix of "wow" and "hmm." The MatFormer2 architecture is the real game-changer here, dynamically switching between models to save battery and avoid overheating. But it’s not perfect. For example, the image recognition nailed my Zermatt photo but missed the exact location, while ChatGPT got it right.
Timestamps:
00:00 - Intro
02:34 - How to install Gemma 3n
Resources mentioned:
• Google AI Edge Gallery APK: github.com/google-ai-edge/gallery?tab=readme-ov-file
• Hugging Face model download: huggingface.co/google/gemma-3n-E4B-it-litert-preview/tree/main
huggingface.co/google/gemma-3n-E2B-it-litert-preview/tree/main
#Gemma3n #OfflineAI #GoogleAI
In this video, you'll learn:
• how to install gemma 3n on android Mobile (even though it's not on the Play Store yet)
• Why this offline AI is different from cloud-dependent tools like ChatGPT or Gemini
• The surprising performance of the E4B and E2B models – including where they shine and where they fall short
I got early access to the developer preview, and let me just say – the results were a mix of "wow" and "hmm." The MatFormer2 architecture is the real game-changer here, dynamically switching between models to save battery and avoid overheating. But it’s not perfect. For example, the image recognition nailed my Zermatt photo but missed the exact location, while ChatGPT got it right.
Timestamps:
00:00 - Intro
02:34 - How to install Gemma 3n
Resources mentioned:
• Google AI Edge Gallery APK: github.com/google-ai-edge/gallery?tab=readme-ov-file
• Hugging Face model download: huggingface.co/google/gemma-3n-E4B-it-litert-preview/tree/main
huggingface.co/google/gemma-3n-E2B-it-litert-preview/tree/main
#Gemma3n #OfflineAI #GoogleAI






![NEW! Wan 2.1 Fun ControlNet ComfyUI Workflow – Bye Bye Glitchy AI Videos
In this step-by-step tutorial, I’ll show you how to use Alibaba’s Wan 2.1 Fun Control Model to turn ordinary videos into eye-catching AI animations. Whether you want to make music videos, social media ads, or cinematic clips, this guide covers everything—from basic setup to advanced stylization tricks.
What You’ll Learn:
✔️ How to download & install Wan 2.1 Fun Control in ComfyUI
✔️ ControlNet techniques for smooth motion transfer (pose, depth & line art)
✔️ Fixing common issues like glitchy frames & style inconsistencies
Resources Mentioned:
• Download Wan 2.1 Fun Control: [Hugging Face Link]
• Free Workflow Templates:
• My Recommended GPU Settings for Lag-Free Rendering
https://comfyuiblog.com/wan-2-1-fun-controlnet-comfyui-workflow/
Join the Discussion!
Have questions about ComfyUI Wan 2.1 or ControlNet workflows? Drop them in the comments below!
👍 Enjoyed This Tutorial?
Please LIKE and SUBSCRIBE for more ComfyUI and AI content!
#ComfyUI #Wan2.1 #ControlNet #AIVideo #ComfyUIWorkflow NEW! Wan 2.1 Fun ControlNet ComfyUI Workflow – Bye Bye Glitchy AI Videos](https://i.ytimg.com/vi/AeaGmjYeMoY/mqdefault.jpg)


![MiniMax H3 Upscaler vs RelayVSR vs SeedVR2 — Which Looks Better?
I upscaled the same MiniMax H3 video with RelayVSR and SeedVR2 in ComfyUI. RelayVSR 2x finished in about 37 seconds. SeedVR2 took 4 minutes 54 seconds.
SeedVR2 gives the sharper result. To me, RelayVSR looks more natural, and it keeps the original face. In this video, I show both results side by side, one problem I found with RelayVSR in fast action scenes, and the settings I use now.
In this video:
- What RelayVSR is, and the 2x and 4x model files
- The two new nodes: RelayVSR Conditioning and VSR Decode
- Why RelayVSR is fast: the big model runs only on every 15th frame and on the last frame
- Test 1: RelayVSR, SeedVR2, and the MiniMax H3 upscaler on one 26.4 second video
- How long each upscaler took
- Test 2: blur and extra detail in a fast fight scene with the 2x model
- The fix: the 4x model with the 4x VAE
- Test 3: the detail slider at 1 and at 0.6
RelayVSR is built into ComfyUI, so you do not need a custom node. If you cannot find the nodes, update ComfyUI.
Settings from the video:
- KSampler: 1 step, CFG 1, no prompt
- Always use the model and the VAE as a pair: 2x model with 2x VAE, 4x model with 4x VAE
- Detail slider: 1 is the default. I keep it from 0.6 to 1, and I do not go below 0.5
- 2x model for normal scenes
- 4x model for fast action. It needs more time and more memory, so use it on small videos
- Detail at 0.6 when the skin looks too sharp
Time in my test (26.4 second video made with [Pika H3 Flex - CONFIRM] on MiniMax H3):
- RelayVSR 2x: about 37.2 seconds
- SeedVR2: about 4 minutes 54 seconds
- MiniMax H3 upscaler: [30 minutes 30 seconds - CONFIRM]
Tested on: [YOUR GPU]. Your times can be different.
Model files for ComfyUI:
https://huggingface.co/Comfy-Org/RelayVSR
ComfyUI/models/diffusion_models: relayvsr_x2_lightx2v_int8_convrot.safetensors, relayvsr_x4_int8_convrot.safetensors
ComfyUI/models/vae: relayvsr_flash_vae_x2_fp32.safetensors, relayvsr_flash_vae_x4_fp32.safetensors
Original model: https://huggingface.co/kopper/RelayVSR
Workflow: https://www.patreon.com/ComfyuiWorkflowBlog/posts/relayvsr-comfyui-175122349
Which result do you prefer: the sharper SeedVR2 or the more natural RelayVSR? Tell me in the comments.
#RelayVSR #ComfyUI #MiniMaxH3 MiniMax H3 Upscaler vs RelayVSR vs SeedVR2 — Which Looks Better?](https://i.ytimg.com/vi/BskTA7BoNsk/mqdefault.jpg)
