3/3/26 UPDATE
I've uploaded a custom merge of the 2603 2-step lora and my key edit of the 2602 8-step lora. This has resulted in a nice, flexible distill that can generate great outputs from 4-10 steps. It has also helped the gridmarking artifacts that were visible at 1.0 strength with the 8-step 2602 UDCAI Edit lora. Please check it out!
Original Post
Credit to alibaba-pai for their fantastic work! This is a step and cfg distill lora to be applied to Z-Image base model and its derivatives. The lora has been converted to be compatible for use in ComfyUI.
Best run at 8 steps with CFG 1. Avoid sageattention. Samplers that add noise per step such as LCM or dpmpp_SDE seem to work well.
Description
A new version update as of 2/10/2026 by alibaba-pai group!
There is now an 8-step version and a 4-step version. Still recommend a strength around 0.7-0.8. Recommended samplers are the same.
FAQ
Comments (39)
Improves the quality of z-image base character loras. Very nice! Even some blurry character loras gets sharper and more flexible with this distill lora!
Glad you are seeing some results!
please share worfklow dude
@vivekkarumudi300 Sure. Here's an image with the workflow: https://civitai.com/posts/26526643
Its a bit of a mess, but it works well with character LoRa's. FaceDetailer takes care of images where the face is smaller and you can lock the seed on 1st and 2nd pass to try different seeds on the face.
Uses the 4-step 2602 version of this LoRa.
For those that generate locally (ComfyUI etc):
You can use Lora with weight 1.0 when shifting the model by 5.0 (e.g. ModelSamplingAuroraFlow). Seems to help with other Loras as well, saying you then can use the usual weights.
Config: DPM++ SDE, simple, 8 Steps, CFG 1
please share worfklow dude
please share worfklow dude
@vivekkarumudi300 Here's an image with workflow: https://civitai.com/posts/26526643
Its a bit of a mess, but it works well with character LoRa's. FaceDetailer takes care of images where the face is smaller and you can lock the seed on 1st and 2nd pass to try different seeds on the face.
Uses the 4-step 2602 version of this LoRa.
Just load the template from Comfyui and follow the steps: "Best run at 8 steps with CFG 1. Avoid sageattention. Samplers that add noise per step such as LCM or dpmpp_SDE seem to work well."
@blackestcurse93
Images hidden due to mature content settings
:(
umm for me the 8 step 2026 version doesn't seem to do anything, I still need high steps to resolve a good image, the original one worked fine.
As you can see in my posts 1, 2, 3, use the 8-Step 2602 UDCAI Edit version @ 0.3 strength. Set z-image base @ 30 steps, 1 cfg, use euler_a/beta.
The quality + speed of this combination is outstanding.
I dunno man, I get 37s on a 4080s for a 1024x1024 image, not 7s. Followed your instruction exactly.
@ynead Your mileage might vary depending on your GPU, of course, but these settings seemed to be the best combination for quality:speed that I found with this lora.
@clueless_engineer What gpu are you using ?
@ynead RTX Pro 6000
@clueless_engineer 10k vs 1k gpu explains the speed difference lol
Question is... why do you use a speed lora if you own a RTX Pro 6000?
@cyberpunk27 If you could gen an image in 7 seconds instead of 30 seconds, wouldn't you rather do it in 7?
@clueless_engineer are you telling me that a 96gb ddr7 gpu takes 30 seconds to gen a 2k zbase image??? You must be doing something wrong.
With ZIB + character LoRA, I get best results with sampler sa_solver_pece and scheduler beta57. It's slower but much better in terms of quality, diversity and consistency. Also, lowering the distilled LoRA to 0.6 and raise the steps to 12, looks like the sweet spot.
For reference, generation time for this setup is about 8.5 seconds (1024x1024 on 5090).
Thanks! Your config works very well. I also changed the shift to 5.
I'm not sure on what settings to use this on, as 8 steps usually creates an image with like 5000% sharpening. I tried other settings like reducing the lora strength, or changing the sampler, but it just sharpens the image to a ridiculous setting. I set the steps to 35 just to counteract the sharpening effect... what good is a 8 step lora if you still need 35 steps lol
When the Lora weight is greater than 0.8, it will cause strange checkerboard textures—they are not obvious, but can be seen upon close inspection.
Indeed I’ve noticed this as well which is why I set the recommended strength as it is. A downside to keeping some of the ffn keys, not observed otherwise.
I’ve already discussed with alibaba-pai and a retrain is in the works.
@UDCAI Thank you for your contribution. ♥
Z-Image-Fun-Lora-Distill-8-Steps-2602_UDCAI works extremely well for me at 0.6 str with res_2s + beta and 8-12 steps
new loras are amazing you manage to make my workflow to 4 steps and the result are fantastic, adherence to my character loras is crazy, where I though I had some drift is now very accurate.
Sorry i'm new here. Is this lora means I can use z image base with fewer steps and lower cfg, so faster generation time like zimage turbo? Do I understand correctly? I have my character lora, can I use mine with this lora together?
Yeah exactly. In my experience these LoRas improves the quality and consistency of ZIB character loras greatly!
@blackestcurse93 thanks bro
@blackestcurse93 Agreed, using ZIB + ZIT together is definitely the new way to go.
Yea you’ve got the idea. It “distills” the model where you no longer need to have cfg (which improved speed alone) but also reduced how many steps needed. The distillation still retains the much broader knowledge of z-image base, so you are not stuck with photorealistic without lora on turbo. Loras also stack much better with zimage base.
With my latest edit, can run anywhere from 4 - 10 steps at 1 cfg.
@UDCAI Thank you for the indepth respone sir. Gonna try this out
@UDCAI ok this thing works like magic. So far I like it a lot. Thank you
Details
Files
Z-Image-Fun-Lora-Distill-4-Steps-2602-ComfyUI.safetensors
Mirrors
Z-Image-Fun-Lora-Distill-4-Steps-2602-ComfyUI.safetensors
Z-Image-Fun-Lora-Distill-4-Steps-2602-ComfyUI.safetensors
Z-Image-Fun-Lora-Distill-4-Steps-2602-ComfyUI.safetensors
Z-Image-Fun-Lora-Distill-4-Steps-2602-ComfyUI.safetensors
Distill-4-Step-Lora2602-b.safetensors
Z-Image-Fun-Lora-Distill-4-Steps-2602-ComfyUI.safetensors
Z-Image-Fun-Lora-Distill-4-Steps-2602-ComfyUI.safetensors