Just messing around with zimage.
I use it like the original either with euler, dpm++ 2s a RF, dpm++ SDE in Forge classic neo, usually 8-12 steps (depends a bit on sampler).
Description
FAQ
Comments (18)
Hi. I downloaded the 11gb model and I took one of the sample pictures to get the workflow. So the workflow loads, I can select my model but when I go to generate I get a "size mismatch for layers.19.attention.out.weigh". I'll try with the smaller mode and see how I go.
Okay I tried with the smaller one too, same error. I should of added the beginning of the error message. See below:
Error(s) in loading state_dict for NextDiT: size mismatch for x_embedder.weight: copying a param with shape torch.Size([3840, 64]) from checkpoint, the shape in current model is torch.Size([2304, 64]).
okay so a comfyUI update has got me past that hurdle. Now I'm onto my next. FYI - I'm using the fallen angel picture workflow. I now get a message at clip node:
ClipSetLastLayer:
nonetype object has no attribute clone
@Unshackled_AI I did not use Comfyui for those images but forge classic neo, so there is no comfyui workflow in those.
@TiwazM okay thanks.
I just checked... the Workflow used in the original photo won't work in ComfyUI because it's missing key elements. I just uploaded a copy/paste of the fallen angel prompt using my ComfyUI workflow. Use that photo to get the Workflow. It will work with any of the Z-Image checkpoints.
PS - The uploaded photo is in the v4 Gallery.
@WillieF That is because I used forge and not comfy ui sorry.
@TiwazM I know. I was just letting him know that when your photo is moved to ComfyUI, it is missing elements Comfy needs to work.
All good. I've actually decided to take a step back from Z-Image Turbo. I'm waiting now for Z-Image Base to be released and then it's gonna get exciting once people start really getting into making the trained checkpoints. I have both SFW and NSFW use cases in mind for me and there are ZIT Lora's that help but I want a trained checkpoint. So I'm gonna take a step back and just give it time. Can't wait to see what stuff gets cooked up.
I will say this. ZIT has been instrumental in producing a picture with specific poses represented in a picture. I've been using that picture then as a depthmap for SDXL to combine with my own SDXL loras. That was fun and really helped.
@Unshackled_AI yes zimage omni should be great as it can apparently do some editing too.
11g模型，8g显存是可以跑的。
trained two Loras? What Loras you merged??
There not example photos here.
2 that I made width ai toolkit, one was real photos and one was nano banana pro
Great work on this one. The checkpoint (fp8) lives up to the z-image TURBO name. Other iterations on the platform seem to not get this. Quality release.
fyi if you re-upload one of the checkpoints as a separate version you wont have the default issue.