Adds turbo back to the base checkpoint, allows you to control the amount of turbo and thus CFG by adding it to base.
This was created by subtracting klein-base from klein and creating a LoRA from it.
Description
FAQ
Comments (57)
4B model in the 9B category. Can you make a 9B version? EDIT: HE FIXED
That would be extremely useful cause, honestly, the 4B model and it's text encoder are very NOT on par with 9B (let alone dev) and it's encoder and it shows in everything, from regular generations to edit tasks, and I can't really comprehend why people even bother with the 4B one when there are plethora of quantized versions of 9B to run on any hardware and those quantized 9B versions will still be quite a bit smarter than 4B ever can be even with tons of loras.
Thanks, I changed it, I'll create one for 9B as well soon
@pirateboogie As I understand it, the speed is faster on 4B vs 9B quant. But we're talking seconds different, so, I don't really get it either.
how are these made? just model lora subtraction?
@ForeverNecessary737716 That's correct, I have a script that does it for me.
@brnfd24434343d 4b can be more gooner friendly in some more fetish zones. Combining that with its faster time, I actually hope for people to do more finetunes on 4b. 9b can fail dramatically compared to 4b in areas involving butts as well.
@Suhnny Bad at butts? 9b has a pretty good understanding of butts, and mixes with many LORA at full strength. Maybe it's a workflow issue.
@Drakeni Workflow issue... Seems like it. There is exactly no reason that 9B would have different butt training data than 4B.
So what does this achieve? Is base with this lora much better than using the turbo checkpoints in general?
we get back negative prompting with the speed of turbo so a win win in every sense , possibly less rigid styles compared to distilled models too but that needs quite a bit of testing to determine, also possible but also needs testing - loras might have better\stronger effect due to the model not being so "set in it's ways" but that's speculation at this point )
Does that mean it can be used as CFG 1 in the base model?
Yes
@anyMODE It works well and results are good. Thank you!
right but then you could also use the turbo directly...yeah? no? i have no clue...
@eyeonyou Yeah you could, but you get less control
Did some testing of 9B rank 128 version - it works very well indeed, thanks for making it ! i also noticed what 9B base + lora actually runs faster than regular distilled 9B even with negative prompt and cfg at 5, but not just that - it looks like it is "smarter" too when it comes to stuff like landscapes\cities\sci-fi etc - regular 9B is definitely not as good in that department - maybe it got a little too overdistilled towards humans generation because bfl tried to make z-image killer (which they totally did))) but overcompensated with that a little, trying to have a good pr this time by appeasing people who howled Flux 2 dev was bad and useless because it took too long generating their only interest in all this ai stuff - them nekkid chicks ofc ))))
This seems to work pretty good! Thanks! If I set the LoRA strength to ~0.5, I can reduce the steps from 40–50 to 10–20, with lower CFG (usually 2.0 - 2.5). It gives me the best of both worlds - the variety and tack-sharp image quality of Base + the aesthetic tuning of the distilled model, all at much quicker generation times.
This is a huge help to CFG
How much control do you want? YES...
Do 9b klein LORA's work in combination with this LORA?
Yes, they should work fine.
"This was created by subtracting klein-base from klein and creating a LoRA from it."
Can "z image base" also be used to create Turbo Lora?
No, I tried, and all I got was noise, the distillation for Z-Image is very different, there are some experiments going on in the community creating a z-image distillation though, so keep an eye out
@anyMODE Thank you for your reply.
I think no because they screwed with the base after releasing ZiT
And we cannot normally distill the base ourselves without its original dataset? Or is it a matter of compute and skill?
@firemanbrakeneck you can distill without the original dataset, you just have to make your own from the original model, it's a matter of compute, there's several ways to do it on consumer hardware, it just takes a lot more time than enterprise hardware. Ostris is working on a ZIB turbo lora at the moment from what I gather.
What did you use to extract it? I tried the easy way using ComfyUI-FluxTrainer, but just got noise. Your result is pretty close to the original
I used my own scripts, I used a randomized energy technique SVD to extract it.
An incredible LoRA, it made my life 500% easier.
🙌💖🙌
The checkpoint I was using insisted on maintaining a saturated color, regardless of the settings, but your LoRA not only sped up image generation, but also gave a magnificent touch to the model's color tones.
Thank you for the beautiful work you did.
any proven good lora weight/cfg/step count combination suggestions?
okay so i used a Lora weight of 1.0, CFG: 1 and Steps: 8 and it gave me real good results
@mrweaz so thats kinda a straight convertion to turbo. I ended up using at 0.25-0.5 with 3.5 cfg 8 steps, speed is acceptable and the quality is much better than in turbo. Turbo edits look like crap honestly.
Any reason to use the rank 64 other than people with low VRAM?
Nope, can't think of any
Just awesome thank you! it works very good with sampler LCM/simple or beta57 even with CFG 3.0. Recommend for you guys to use Flux.2 KLEIN enchancer custom node.
works very well, not good. Good is an adjective. Well is an adverb. Works/working is a verb. Nothing can ever 'work good'. Something can 'be good' though.
works very well, not good. Good is an adjective. Well is an adverb. Works/working is a verb. Nothing can ever 'work good'. Something can 'be good' though.
Nailed it. Took it apart, rebuilt it properly, wired some of the logic from Qwen — now every run just like magic. I will share my workflow after I clean all private stuff. but its gold
Can we definitely use negative prompts with this BFL says that Klein doesn't support them
It would be good to have a short description how you suggest to use it properly. Number of step counts, samplers, difference between ranks, CFG, strength etc.
I think it's 8 steps, Sampler: Euler, CFG: 1.
Just want to say: extraordinary! Negative prompt AND fast generation!? And it works great. Thank you so much! Good work!
What settings are people using for this? At full strength, and at partial strength. There's no info in this description.
There's no info because I had no clue myself. I put it out there for others to find out.
I personally would try at 0.7 strength, 8 steps, cfg 1.5 and work your way down in strength and up in steps and cfg from there to find a nice compromise
Intensity 0.4-0.6, raw images CFG3.5 with 10-12 sampling steps, image editing CFG1 with 8-10 sampling steps. If too many anatomical errors are encountered, the number of sampling steps can be increased.
probably a noob question but whats the difference between the 9b 64 and 9b 128?
The size, that's about it
@anyMODE hm ok. so it doesn't make a difference which one i use then? thanks. its been working great for me
Rank 128 delivers better quality and stronger effects in most cases, but costs more memory and loading time. Rank 64 is the more resource-efficient version and is often sufficient when you want to save VRAM or test quickly.
My Lora Manager can’t detect the Klein 4B/9B Base to Turbo, it detects other Loras in my Lora folder fine but not this one, anyone else having the same issue?
I had the same problem. This fixed it:
1. Open Lora Manager (square blue button in top right toolbar, left of the Comfy Manager button)
2. Click dropdown arrow next to the Refresh button, then select Rebuild Cache
3. If that doesn't fit it, check your loras folder for a .metadata.json file with the same name as this lora. Delete it.
4. Rebuild cache again.
Those steps fixed two separate issues I had with loras not showing up. I did throw in a few Ctrl+Shift+R's to refresh the page cache too, so you could try that if the above doesn't work.
noob question, why and when do use this? I tried plugging it in to a workflow and the results looks like im on lsd lol
It's a LoRa for faster generation (few steps) on a 9b base model. Useful if 9b base is provided, but not a 9b distilled version. You will get terrible results if you use it on a distilled version.
@aising23 Ah I see thanks, yeah I don't use base version but maybe I'll try it.
I've noticed that using this Lora seems to reduce editing strength – sometimes the image shows no change at all compared to not using it.
I noticed that when CFG is set to 1, the first step follows the prompt correctly, but from the second step onward, it reverts back to the original image. Only when I increase CFG to 1.2 does the editing work properly — though at the cost of roughly double the computation time.
Just wanted to flag this. Thanks!