⚙️ Key Features of this Model:
Advanced Merge Methods:
Uses techniques like weighted sum, interpolation, and selective update to deliver efficient and high-quality outputs.
Layer Importance Optimization:
Early layers are optimized for speed, while the later layers are designed to enhance image quality.
Dynamic Tensor Resizing:
Ensures seamless compatibility between different model tensors, making your setup flexible and robust.
****NF4 version****
🎛️ Setting Up GGUF Support:
-FOR WebUI Forge is installed and properly set up.
https://civarchive.com/articles/6715
🛠️ Complete Installation Guide
📁 Setup Structure
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 diffusion_models/
│ │ ├── (basic)📄 bernoulli.gguf
│ ├── 📂 text_encoders/
│ │ ├── (basic)📄 clip_l.safetensors
│ │ ├── (option1)📄 t5xxl_fp16.safetensors
│ │ ├── (option2)📄 t5xxl_fp8_e4m3fn.safetensors
│ │ └── (option3)📄 t5xxl_fp8_e4m3fn_scaled.safetensors
│ ├── 📂 vae/
│ │ └── 📄 ae.safetensors
💎 Essential Components
This merged model offers a balanced solution for AI-driven image generation, emphasizing both speed and quality. Whether you're processing single images or large batches, it delivers high-quality visuals efficiently.
🔤 Text Encoders - The Brain Behind Natural Language Understanding
Note: You only need to choose ONE of the T5XXL options below based on your hardware capabilities
T5XXL Options (choose only one):
🎭 VAE - The Visual Artist
👨💻 Developer Information
This workflow guide was created by Abdallah Al-Swaiti:
For additional tools and updates, check out the OllamaGemini Node: GitHub Repository
Description
best of 4th generation
FAQ
Comments (19)
Hello my friend wish one is better q8 or bf16?
bf16 more quality more slow , its depend on gpu if 24 go with bf16 if 12 gpu go with q8 , i favourite q8 and q5km witch is faster
@AbdallahAlswa80 Q5_K_M is better than city96's Q5_K_S? Q5_K_S is my favorite version and it many times beat Q8 with the same seed, imho, even if it makes no technical sense. That version's output is just great. I have to test your version.
@AbdallahAlswa80 thank you, also in comfyui some some of your model gave me an issue I google it a bit and i read sometime about that need a RTX to be able to run it but I have an old NVIDIA M40 24gb obviously not an RTX so wish model you can recommend me. I really like and appreciate all the work your doing here.
@Artikuz what the error text do you have
@AbdallahAlswa80 the error is this one Error occurred when executing SamplerCustomAdvanced: module 'gguf.quants' has no attribute 'dequantize'. the issues log is really long. If you need the complete log let me know and I send it to you.
@Artikuz okay i think this issue refer to not update the gguf model loader
https://github.com/city96/ComfyUI-GGUF
1- update to latest
2- pip install --upgrade gguf
3- be sure that u downgrade pip install numpy==1.26.4
@AbdallahAlswa80 I think the model doesn't run because is a BF16 the other models I down load from your profile work just fine, my gpu can do fp16 and fp 32 but maybe cant do bf16.
@Artikuz i read about that before some cards not supported bf16 buts its okay
@AbdallahAlswa80 yeah is ok, I be using your other checkpoints versions instead also thanks for the assistance.
'bernoulli_q5KM.zip" has the gguf file buried in 11 directories deep, including your operating system's username.
HHHH ,,, I USED CMD to COMPRESS THAT , thank you
I assume you're waiting to update the Q5 file, currently listed as missing xD
@MysticDaedra its updated
please just give me like if it better than dev
edit: Okay, so you actually did new versions: Q3_K_L, Q4_K_M and Q5_K_M. And they are different than what city96 did. So Q6_K and Q_8 are your remakes?
I downloaded Q6_K for testing. Is it better than Dev? Sure. Faster than city96's Q6_K? About the same speed with my rtx 3060. I'm curious to know: These versions are merges of what? What exactly is the justification for these versions? I hardly see any difference. Slightly different output than city96's Q6_K? Yes. But is it better or worse, I can't really say.
this is my personal merge i used coding for this , and the python script still under construction , i was did merged before using dev plus schnell in this model https://civitai.com/models/661102 , but bernouli is diffrent it has difference features (the structure and the origins are different) , please more feed back about your experience
i,ll appreciate your help of comparison , i have to much pressure here
We need more details about each version, comparison, differences..
If possible any recommendation for settings, prompting..etc
Thank you.
as you go 3qkl ---------------------> fp16 need more memory , slower , more quality , the steps 3-6 , cfg 1 did you read the description