Simple Z Image Turbo Workflow
With the batch workflow, you can increase the inference speed per image, the bigger the batch, the more speed/image you get.
Description
v1.05
Add LoRA support https://github.com/comfyanonymous/ComfyUI/pull/10978/commits/52279fc86beb8e3439ec413e3f2aa28c04623c5e
v1.04
Add Z_IMAGE_TURBO_BATCH_Variant
v1.03
Add current used prompt list for "Z_IMAGE_TURBO_BATCH", as a string output
FAQ
Comments (56)
face detailer maybe?
Well we can add alot of things ,) But the first one gotta be simple
I will add a new one for batch processing, to avoid the ping pong loading/unloading the text encoder after each renders, this will be much faster to render multiple images with different prompts.
where is the workflow? the zip file contains picture?
drag one into comfyui
@surk omg ty. I got it
error message below
CLIPLoader
Error(s) in loading state_dict for Llama2: size mismatch for model.embed_tokens.weight: copying a param with shape torch.Size([151936, 2560]) from checkpoint, the shape in current model is torch.Size([128256, 4096]). size mismatch for model.layers.0.self_attn.q_proj.weight: copying a param with shape torch.Size([4096, 2560])
everything is the same like in original workflow. How do I fix this?
It seems for this model to work, you need to update comfy. If using portable, there is a bat file for updating or you can use the manager menu to update. I use the non-portable locally and to update you click the comfy icon > help > check for updates and it will install. Hope that helps anyone else running into this
me too
Uptadate comfi. It fix this issue.
I don't get it, how do I change the model name? I have the turbo model on civitai downloaded and can't use it?
What you mean by can't use it? Just select your model and run it
@pgc It seems I cannot link the model folder from forge into comfy as unet. Tried putting z-image into comfy itself appears to work
@cliang96844 @cliang96844 use this https://github.com/arnobpl/SymlinkCreator and create link of z_model file from forge to comfy's (to keep free space and not duplicate models).
you can clone node then everything ok
I'm not seeing how this is any better than ComfyUI example workflow. The outputs are exactly the same, and none of them look remotely like your example images.
I don't know what you're expecting, is it advertised like something that would make the output looking better than a regular example workflow? No, it's not the point. It looks good, it's simple and fast, doesn't need anything else.
As for your assumptions that i'm cheating with the images I posted, even if I posted a video of the workflow in use, that is funny work from you
People like Joey here are the reason why a lot of creators put their stuff behind a paywalls. The sense of entitlement of people in this community sometimes is unbelievable.
Uhm.... the zip file only contains 2 png files, no json
It's the same, drag and drop into comfyUI
Наш слон!
How to set up batch generating? I have image.txt with prompts one by one under each other, each from ne paragraph, so I copy the breadcrumb (C;\newfolder\image.txt) in your WF but It doesn't work ;/
Huge fast!
How does your 'Z Image Turbo Load Models' have upscaler in it? it isnt in the wf.
You mean the upscaler link? You can ignore it, just forgot to delete it
@pgc Sadge. Been trying to learn how to upscale on comfy, and no wf tends to have a functional one.
@SomebodyOnceToldMeh This workflow initially had Ultimate SD Upscale as a second pass, but after some comparisons, I've seen that it was not worth the time consumed.
You can just render at higher resolution, 1920x1080 or vertical would work without problems. You don't really need to upscale post latent decode
I'm a noob...
I can't find these knots anywhere.
Can someone help me here or privately, please?
knots? you're talking about the nodes? Well if you have comfyUI manager, the nodes will be shown as "missing", you click on Install missing nodes, restart comfy and it's done
@pgc Bro, forgive my profound ignorance, but where is the node called "Z image Turbo Inference" located? Where can I download it? For you experts it's all easy, I understand, but for us noobs it's a labyrinth with no beginning and no end.
@llmsonoio1234356 I see, the "Z image Turbo Inference" is not a node by itself, it's a subgraph, see it as a simplified group of nodes.
When you select multiple nodes, you can convert them as a subgraph to make things cleaner, that's what I did here. To recognise subgraphs, you can see a gray square on the top right of the node, when you click on it, the subgraph will opens and you will see what it contains.
Since I don't really know if you have an actual issue or not I won't get further in the explaination,
Do you have an issue where the nodes are not being installed? Do you have errors? Please provide more informations on your current issue, so It will be more covenient for me to explain what is going on, and what you can do from this point
solved! It was a ComfyUI update issue.
what node pack gives the execution timer?
Mine, search for "fancy" in comfyUI https://github.com/PGCRT/CRT-Nodes
@pgc thank you, sorry for the dumb comment, i found it shortly after posting lol
@SeoulSeeker No problems
Error:
KSampler
No module named 'sageattention'
Set sageattention to "disable", it should be set to auto by default, wich would disable it if you don't have sageattention installed
@pgc ok thank you. Do I need it though? What is it if you don't mind?
Redditor made a easy install to install sageattention on to your system here a link: https://github.com/Justify87/Install-SageAttention-Windows-Comfyui
if you need help i can guide you
Are you able to explain or add to the workflow where does one put a lora loader, and which one to use?
Unpack and expand all nodes
@uqw I think I'm doing something wrong then, I did that and don't see lora anywhere...
@droegan LoRa was still not compatible last time I updated the workflows, so there is no lora loaders, you can just connect a lora loader before the sampler, I will update it, probably later i'm abit busy rn
I use LoRAs. ComfyUI gives me errors, but LoRAs definitely working.
Does this work if you replace with GGUF or AIO model?
this line on your reddit post
random-prompts directory (about 2.6k prompts) included in my nodes toolset
............................ can we use this with anything like sdxl, illus, pony
............................ do you have a github with this nodeset?
Yes you can use them with any diffusion model you want, if you already have my nodes set you can find this one "Image Loader Crawl (CRT)" that one will crawl to your prompts folder, I use it alot.
This is the directory of the prompts, you probably already have it under "./ComfyUI/custom_nodes/crt-nodes/RandomPrompts"
https://github.com/PGCRT/CRT-Nodes/tree/main/RandomPrompts
@pgc okay cool super cool, i don't have it yet, just starred your node set and followed you on github : )
thank you for this, i can tell it great : ) this setup is going to be fun with challenges on here once i get my todolist caught up
I'm getting an error after the latest ComfyUI update: CRT_KSamplerBatch
Pooled_output
Thanks for the report, yeah they tend to broke things that works already,
Try to update your CRT-Nodes suit and it should be good, tell me if you have any other issues, thanks again
getting an error: Error while deserializing header: header too small
00:47:816 per picture -32 gb Ram 16 gb Video.
This not Turbo
~3:00 or 180s on a RTX 3080TI 12GB | 10 Photos @ 960p w/ 2x Upscaler. Using the Qwen3 4b Q5_0 and ZImageTurbo_Q5_K_M.
Pretty good way to automate multiple generations. Nice