3bQ8 vs 7bQ8 vs 7bQ5 comparison

[naxci1]

3bQ8 - 7bQ8 - 7bQ5

---

---

[naxci1]

3bQ8:
tensor_count = 635
Pointer size: 135 Bytes
Size of remote file: 3.66 GB

---

7bQ8:
tensor_count = 1128
Pointer size: 135 Bytes
Size of remote file: 8.84 GB

---

7bQ5:
tensor_count = 1128
Pointer size: 135 Bytes
Size of remote file: 5.78 GB

---

Question: How come the smaller model with fewer tensors produces higher quality images than the others?

[adrientoupet]

Hi @naxci1 - Are these results still true when using the non GGUF version ?
If you can share your input frame (even just a frame, not the video) so we can reproduce it, that would be handy. Thanks

[naxci1]

---

---

---

[naxci1]

Hi @adrientoupet

I think the best ones are 3bfp16 and 3bQ8 models.

The 7b models, although larger, never gave me good results.

[Fenixdg3duy]

test.mp4

Using seedvr2_ema_7b_sharp_fp16.safetensors. 640x360 base video resolution, batch 125 frames x4 multiplied resolution.
Using the naxci1 workflow on an NVIDIA RTX PRO 6000 Blackwell 96GB card.
I'm sorry to raise more doubts, but if it helps, I'm experiencing the same issue as naxci1. I always used 3bQ8 until yesterday, 12/8/25, when I was able to use seedvr2_ema_7b_sharp_fp16.safetensors, and the result was more than optimal.

[naxci1]

Hi @Fenixdg3duy

NVIDIA RTX PRO 6000 Blackwell 96GB - are you using this on the cloud? Which site are you using it on and is there a ready-made template?

[Fenixdg3duy]

I will explain the steps to use ComfyUI in RunPod with a template:
NVIDIA RTX PRO 6000 Blackwell 96GB

Basic steps:

1. Select the Template

Go to RunPod and in the "Community Cloud" or "Secure Cloud" section

Look for a ComfyUI template (there are several popular ones; lately I have been using one called Blackwell compatible, or else the template that is the latest version of comfyui)
Check the specifications: CUDA version, PyTorch, pre-installed nodes

2. Configure the Pod

Select the GPU you need (in your case, you have been working with Blackwell architecture)
Configure the storage (the basic one is sufficient)
Set the maximum budget; a one-time purchase of $10 for testing is more than enough.

3. Start and Connect

Once the pod has started, you will see a "Connect" button
Use "Connect to HTTP Service [Port 8188]" to open ComfyUI in your browser
Or use "Connect via SSH" if you need terminal access

4. Upload your Workflows

Upload your workflow .json files from the ComfyUI interface
Or copy files via SSH/web terminal

5. Manage Models

Models are stored in /workspace/ComfyUI/models/
You can upload them manually or use the terminal to download them with wget
Make sure to place them in the correct folder (checkpoints, vae, loras, etc.)

6. Install Custom Nodes

From ComfyUI Manager (if installed)
Or manually via SSH or manually in the console option, which is like CMD or Windows Terminal.

[proxyid]

Same for me. 3B models are crisp and sharp, the 7B models are blurry. And the 7B_sharp models are everything else than sharp.
Has been like this for me since 2.5, maybe even in the old nightly, I cannot remember.

[disk02]

I think of it in terms of cfg. The higher the cfg the more the output will look like the input, and a lower cfg gives the model room to be more creative. A lower cfg can be useful if an input is really low quality or needs restoration. While a high quality input, where there are fine details that you would like to preserve would call for a higher cfg. I would say the 3b model has the lowest cfg, then the 7b_sharp model sits in the middle, with the 7b model having the highest cfg.

[proxyid]

But it wasn't always like that. And as far as I know, the models did not change?

[Junchien999]

The 3B model is quite helpful for blurry images. I tested it on 200-300 pictures. When the current image is blurry (though this concept of "blurriness" is hard to describe, I can clearly distinguish it myself), the hallucination rate of the 3B model is quite high. You can detect this through facial comparisons. You could try using 3B or FlashVSR to reduce blur first, and then apply the 7B model, though this approach is quite costly. Still, it can yield excellent results.

[proxyid]

What upscaling factor do you use for the "first" path? Or do you upscale and downscale again?

[Junchien999]

你用什么升频因子来做“第一条”路径？还是你会先升级再缩放？

The original image is processed through the 3B-F16 model, and the resulting output is then downscaled to a lower resolution before being fed into the 7B-F16 model.

[naxci1]

你用什么升频因子来做“第一条”路径？还是你会先升级再缩放？

The original image is processed through the 3B-F16 model, and the resulting output is then downscaled to a lower resolution before being fed into the 7B-F16 model.

I got the best results with the 3BQ8 model and I always use it.

[Fenixdg3duy]

@Junchien999 junch
What an interesting technique! First, you use the low-quality 3B version, and then you apply 7B to the result. I like it. I did the corresponding test, but I can't see any details in the hairstyle or hair accessories. Perhaps the initial resolution I used in 7B is low in the image, less than 1000 x 1000. What do you recommend to gain detail in the hairstyle and accessories?

[Fenixdg3duy]

@Junchien999 I would like you to share your configuration, as I can't seem to achieve the level of detail I see in the image, no matter what I try. The accessories and hair look great. Can you share your .json file? I'm dying of curiosity.

[zelenooki87]

@naxci1, have you changed your workflow in the meantime? Do you still mostly use the 3bq8 GGUF model? I noticed your repo-fork for flashvsr, and it works but unfortunately the results aren't as good as with seedvr2. Can you tell me what the best settings are for the torch compile node? I have all the necessary libraries, I'm wondering if it only compiles once and does it always have to be in the workflow for the model to stay optimized when loading? Also, does torch compile work for GGUF or only for PyTorch models? Thanks

[naxci1]

Hi @zelenooki87

I'm currently using this workflow; it's easy to do both automatic and manual input. I always use the 3bQ8 model, as I get the best quality with it. My videos are in extremely poor condition; others don't give good results. Flashvsr isn't very high quality, so I improved it to a new version, both for experience and to add new VAE models. For videos in poor condition, the SeedVR2 model is undeniably the world's number one, but if the videos are in medium or good condition, Flashvsr is possible. It's actually faster and gives better results than Topaz. Its training model isn't good, that's why it doesn't give quality output. The SeedVR2 training model is really good, even better than Topaz SLM and SLS. The only problem is that the VAE is slow, but they've created such a DNA that you can't add other VAEs. I spent weeks trying to add many fast VAEs, but because the DNA didn't match, they either didn't work or gave poor quality output. But with Flashvsr, I added 5 different VAE models in one day, and they all worked very well. We're expecting the SeedVR3 model this year; the developer promised to release the new version this year, and we're waiting for it. It would be great if VAE model selection was also included. The DIT model processes quickly; the only slow part is the VAE decode section. I tried very hard with Torch and SA3, they worked, but I gave up because the effect wasn't good. They work well on Linux, but I couldn't get them to work quickly on Windows. Torch works on all of them, but you need Linux. If you have Linux, there are tutorial videos on YouTube for installation.

[zelenooki87]

@naxci1 Awesome to hear about the SeedVR3 model! :)
It will really hit the spot if it's out-of-the-box compatible with the current ComfyUI-CLI implementation.
All the work on the ComfyUI repo is really paying off now. :)

[skv89]

Hi @zelenooki87

I'm currently using this workflow; it's easy to do both automatic and manual input. I always use the 3bQ8 model, as I get the best quality with it. My videos are in extremely poor condition; others don't give good results. Flashvsr isn't very high quality, so I improved it to a new version, both for experience and to add new VAE models. For videos in poor condition, the SeedVR2 model is undeniably the world's number one, but if the videos are in medium or good condition, Flashvsr is possible. It's actually faster and gives better results than Topaz. Its training model isn't good, that's why it doesn't give quality output. The SeedVR2 training model is really good, even better than Topaz SLM and SLS. The only problem is that the VAE is slow, but they've created such a DNA that you can't add other VAEs. I spent weeks trying to add many fast VAEs, but because the DNA didn't match, they either didn't work or gave poor quality output. But with Flashvsr, I added 5 different VAE models in one day, and they all worked very well. We're expecting the SeedVR3 model this year; the developer promised to release the new version this year, and we're waiting for it. It would be great if VAE model selection was also included. The DIT model processes quickly; the only slow part is the VAE decode section. I tried very hard with Torch and SA3, they worked, but I gave up because the effect wasn't good. They work well on Linux, but I couldn't get them to work quickly on Windows. Torch works on all of them, but you need Linux. If you have Linux, there are tutorial videos on YouTube for installation.

How come you disabled the swap_io_components, offload_device, and cache_model?

[naxci1]

Hi @skv89

It's better to disable swap_io_components.

I'm not disabling offload_device; selecting none ensures they are stored in VRAM, and it provides information when you hover over them in the workflow. Processing is faster with VRAM models.

If we store the models in VRAM, i.e., select none, then cache_model = false must be set. Otherwise, it gives an error because it can't cache in VRAM.

[skv89]

Also for attention_mode, how did you install sageattn_2? Does it make any difference in speed? And were you ever able to get the Torch_Compile node working after the latest updates?

[naxci1]

Hi @skv89

The link below contains all the ready-made Windows AI Wheels; just download and install the correct one according to your Python, PyTorch, and CUDA versions. You can get help from AI; installation is very simple. Currently, SA2 is the fastest on Windows; I installed SA3, but it doesn't support GPUs and requires an H100 to work. I've actually been working on Torch.compile for two days now; it worked, but compiling C++ is extremely slow on Windows, but it does it very quickly on Linux. I'm currently testing it, and I've spent two days on these tests. I still can't get the desired results in the Windows environment. Therefore, don't use it on Windows for now; use SA2.

https://github.com/naxci1/AI-windows-whl