This is a refined version of the original Krea 2 Turbo Distill LoRA from TheDivergentAI.
Base Model: Use with Krea 2 base model only.
What I did:
Removed unnecessary blocks to make the file smaller
Adjusted remaining blocks to keep image quality high
Allows the base model to have more control over results
Faster loading and generation
Original LoRA by: TheDivergentAI
Original source: https://huggingface.co/TheDivergentAI/krea2-turbo-distill-lora
Note: This is experimental. Test it yourself to see if you like the results.
Examples included: Original vs Refined comparison images with full metadata and workflows in the ZIP file.
Workflow: https://civarchive.com/models/2608465/qwen-and-krea-2-sam3-face-inpainting-workflow
Removed Blocks & Their Functions:
other_weights - General model adjustments and miscellaneous parameters
first - Initial layer processing at the start of generation
last - Final layer processing at the end of generation
txtfusion - Text fusion features for prompt blending
block_1, 2, 3, 4, 7, 8, 9, 11, 12 - Lower-impact transformer blocks with minimal contribution to final output
Kept: 19 most important blocks (block_0, 5, 6, 10, 13-27) that provide the strongest style and quality improvements.
Version 1.1 - What's Improved:
✅ Dynamic Angles Restored
Found and adjusted the blocks responsible for varied, interesting compositions.
No more perfectly symmetrical. Images now have natural, dynamic perspectives.
✅ More Details & Sharpness
Bumped up values in key blocks that handle fine textures.
Images are sharper and clearer.
✅ Smaller & Faster
Trimmed down to just 13 essential blocks (from 31).
Faster loading and generation, smaller file size.
What I did:
I tested different block combinations to find which ones control angles and details. Then, I slightly increased their values to bring back the quality while keeping the file lean.
Important Note on Sample Images:
The sample images here were generated using only the Rank 256 Version 1.1 LoRA. If you want to compare them with the original or Version 1, please check the sample images on the V1 page.
Note: Still experimental - test it and see if you like the results!
Description
krea2_turbo_distill_r256_Refined, krea2_turbo_distill_r128_Refined
FAQ
Comments (12)
sorry i'm a noob but what exactly is the purpose of this lora? makes the generation faster?
You can use raw model with 8 steps (faster) gives different output than turbo model
@xion1 thank you so much
do you have instructions on how to use it?
steps 8, cfg = 1.0, lora weight 1.0, krea raw model
Thanks for this. From your samples, there is definitely quality loss: For bubble-lady and the dragonfly, the camera shifts to a less interesting angle: it's perfectly on axis and symmetrical. The market scene has less variety: in particular the sacks on the left shift to complete uniformity in size, height, and position. The underwater lady shifts to a less interesting, more symmetrical composition in the background. I mention this in case it's fixable. For some, the pros may outweigh the cons.
Thanks for the detailed feedback. You are absolutely right. Looking at the images, removing lower impact blocks caused the compositions to become too symmetrical and uniform (like the perfectly centered dragonfly and identical market sacks). Those blocks were clearly responsible for dynamic angles and natural variety.
I will try to adjust those blocks to fix the composition issues and upload an updated version soon.
@eagle42 separately from turbo-ization, If you've discovered blocks that are responsible for dynamic angles and variety, and if it's possible to adjust them in order to increase that effect alone, I think that would be an extremely popular lora!
@terrariyum I am working on it. The problem is testing different block combinations and values takes a lot of time and patience. There are also other factors like seed, prompt, and steps to consider. For the turbo effect, you don't need all layers I have reduced it to 12 blocks and still get good results. I am still hunting for the specific block responsible for dynamic angles.
Hi, since you seem to have a lot of knowledge on this stuff, have you heard about Mr Flow aka Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling?
It gives a MASSIVE boost in performance to all AI models, like 10x speed boost, (21 times boost with Z image turbo, 25x for Qwen) but its not been converted for Krea2 use... I dont know if you have the skills to do it... but anyway, more info here :
https://github.com/Xingyu-Zheng/MrFlow
I just thought you would find this very interesting... and maybe you can do something with it for the benefit of everyone.
Anyway, great work and thanks, hope this helps.
RealRebelAi ported for its Krea-2 support @ https://github.com/RealRebelAI/Rebels_MrFlow
I am not an expert, just a hobbyist trying to learn how these models work.
Thanks for the MrFlow link. I checked it out. It creates a low-resolution image first, then upscales and refines it. This helps keep the image quality while also making the process much faster than a normal latent upscale.
Thanks for sharing it. I am definitely going to add it to my workflow and try.
I also saw lokalfokal's reply that RealRebelAI has already ported it for Krea2.