The latest version workflow and nodes are here
https://github.com/peterkickasspeter-civit/ComfyUI-ZImageTurboProgressiveLockedUpscale/tree/experiment
=====
Custom node: https://github.com/peterkickasspeter-civit/ComfyUI-ZImageTurboProgressiveLockedUpscale
Workflow is in the showcase images. More details here - https://www.reddit.com/r/StableDiffusion/comments/1r1qyj2/zimageturboprogressivelockedupscale_works_with_z/
Distill lora for base - https://huggingface.co/alibaba-pai/Z-Image-Fun-Lora-Distill/tree/main
*use either Z-Image-Fun-Lora-Distill-8-Steps.safetensors or Z-Image-Fun-Lora-Distill-4-Steps-2602-ComfyUI.safetensors
Description
FAQ
Comments (40)
Not seeing prompts or generation info on the sample images.
ouch. pls refresh now. its there
I generate images at 1152 (width), 896 (height) and in this Lora unfortunately they are very blurry. I use ComfyUI and flux1-dev.safetensors checkpoint
try euler and beta. see this post - https://github.com/lllyasviel/stable-diffusion-webui-forge/issues/1157
@peterkickasspeter it helped, thanks!
Hi, I want to try this out but getting error in comfyui specifically with this lora only. There are many lines like this:
lora key not loaded: lora_unet_down_blocks_0_attentions_0_proj_in.alpha
...etc
Any idea?
Further: Did you upload the wrong lora? Perhaps this one is actually sd / sdxl? Because Flux doesn't have a unet, it uses a transformer. The sd models use unet
@nyeb5270131 which checkpoint are you using?
@peterkickasspeter The one here, https://huggingface.co/black-forest-labs/FLUX.1-dev/resolve/main/flux1-dev.safetensors
if you are using the original checkpoint, you should not have issues. maybe you should update your comfyui or forge
@peterkickasspeter Think it's working now, might have been due to a custom node I had installed. Clean comfy install works thanks
This lora weirdly turns most people into chubby versions of themselves lol. It would be great if it focused on the image quality and not body types :D
hmm okay. i will see if i can adjust the dataset in the next version. thanks for the feedback
great feedback. I experienced the same with young females.
@peterkickasspeter I'm wondering if it's just a captioning issue. Does the captioning include descriptions of the people in the image? e.g. if they are chubby, it likely needs to be captioned as chubby so it ignores chubbyness as trait of the lora's style
I agree. Great LoRA. But instead of changing just the image quality, it changes the content to a large degree.
@peterkickasspeter Yeah, if you don't caption the people's actual appearances (e.g. weight, height, sex, etc.) you are training the lora to apply what is in the sources to what you are captioning. So if you caption "photography of three women" and they are all overweight, the lora is learning that "women" is the overweight subjects in your dataset.
Lora is cool, seems to have trouble with body figure / anatomy in lot of cases, similar to sd3 stuff. Any way to improve? Thank you
try to generate at higher resolutions. the anatomy is mostly good. flux is able to handle any resolution as long as you have enough vram/ram and tiled vae
@peterkickasspeter What res do you use? I used around 768x1344 before
I constantly getting an error when generating images. It seems like it doesn't work at all unless you have a rtx 4090...
i can generate fine on my 4060ti 16gb and 32gb ram. maybe you should update your comfyui or use the quantized models (if you go below q8, the lora is doing minimal changes)
@peterkickasspeter I'm using Forge and I tried every Flux Dev model but I constantly get the same error. Maybe my rtx 3080 is too weak
@peterkickasspeter I apologize. I just found out that no Lora works with my flux, no matter what Flux Lora I use it gives me an error so it's not your Loras fault.
Best real lora. I use Forge
me too and it's really the most useful lora on CivitAI. by far !
I wrote an article about style adherence referencing your LoRA: https://civitai.com/articles/6835
Thanks. I see some issues with v1, like you and others pointed out. v2 is cooking. I think it will fix all those issues. Please stay tuned
@peterkickasspeter Will do, thank you
@defnotarobot I wait for your feedback on v2 :D
This doesn't seem to work in Forge. When you click the Lora it doesn't add it or apply it to the prompt.
same here
@dmovalli Glad to hear I'm not going crazy.
Try the new version. That one works for me.
it should work. i test everything in forge. try to update it
Tried generating in Forge on Q8 and Lora models. It takes 3-4 minutes to load each new Lora generation and then another 3 minutes to render. For a very long time.
setup? If you have 12gb vram (or less) use the q4_K_S or Q6_K
It almost doesn't matter what the settings are, many LoRas load the generation, not only that, so here we should expect some new updates and techniques, it is clear that on powerful video cards it will be much faster.
@downwego RTX 3060 12GB FOrge
Definitely adds an amateur quality and is a great start! A few notes:
It does seem to degrade image coherency a bit (note, this is different from image quality, which we want to be somewhat bad for this LoRA)
It does seem to cause Flux to struggle with certain subjects/concepts. For example, this LoRA is extremely biased toward white people, at least when it comes to male subjects. Prompting for an African American male fails most of the time when using the Lora, but doesn't fail when just using Flux alone.
I imagine that as more iterations of this and other LoRAs are created, we'll see some of these issues ironed out, but wanted to provide the constructive feedback.
Agreed. I am trying to fix them in v2
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.