v18 Feedback
Hi I couldn't find a thread yet about v18 specifically so I thought I'd post one.
Just wanted to say fantastic job on v18. No stripes or grid issue that I can see. Prompt adherence excellent, Character consistency also excellent. All in all - excellent. I was previously sticking to v16, but now v18 is what I'll be using.
One of best as I have seen in Qwen models.Congrats.
In the new image, the new image is offset upward than the original image. How to solve the offset problem?
In the new image, the new image is offset upward than the original image. How to solve the offset problem?
Are you using my fixed text encode node?
我也做了一些v18的测试，虽然只是初步测试，但是效果非常的好，可以说是10.4之后最好的版本
For me I feel v16 is still the best for edit consistency more than v18 (from my testing editing two characters in a image). The faces were not consistent with v18. But when i tried text to Image, v18 felt slightly better than v16 especially with NSFW positions.
But for pure NSFW text to image I still keep going back to v11.4 and really love it
What is the best prompt to do face swap? I felt the model could do face swap, but the ratio is off (the head is too big) and didn't follow the source expression.
条纹还是很明显的。
Why is the image I generated like this? Was the model file corrupted when I downloaded it?
Why is the image I generated like this? Was the model file corrupted when I downloaded it?
Interestingly as of yesterday morning I've been having this same issue. And it's not just with this model, but with other Qwen models I've tried. For example I tried the default workflow for Qwen-Image t2i, same result. Although interestingly this happens less often compared to i2i. I'm not sure what has caused it. I've tried downgrading ComfyUI, fresh installation of ComfyUI container, all sorts. No luck. This randomly started happening while generating images yesterday morning, suddenly this noise. With t2i workflow I'm at least able to get a proper image output every 1 in 4 generations or so, the rest have noise or glitches.
If you find the solution I'm all ears, I'll also post if I find a solution, but so far I'm stumped. I also asked on the ComfyUI Discord server, one person replied but it wasn't a solution, so no help there. Given it's not likely my GPU is broken (RTX 5090), at least some simple math tests passed with expected results, At the moment I can only assume that it's either something in ComfyUI or something that ComfyUI is using. I've even tried downgrading torch.
The V18 is really good
Why is the image I generated like this? Was the model file corrupted when I downloaded it?
Hi @oumatsu - not sure if it's helpful but I've resolved my problem which I believe was similar to yours. Turns out my GPU might be suffering slight degradation even though it's had light usage and isn't by any means old.
Anyway, to solve my problem I ended up reducing the clock speed offsets for core and memory and reducing power limit to 69% in MSI Afterburner while testing on my Windows 10 virtual machine. Since doing that I've not had an issue so far. I've inspected the 12 pin connector on my MSI Ventus RTX 5090 GPU but I can't see any issues there, so I assume unfortunately either (more likely) the GPU has had a slight bit of degradation, which isn't what I'd expect at the moment. Anyway hope that information helps, unless your issue is caused by something else.
Why is the image I generated like this? Was the model file corrupted when I downloaded it?
Interestingly as of yesterday morning I've been having this same issue. And it's not just with this model, but with other Qwen models I've tried. For example I tried the default workflow for Qwen-Image t2i, same result. Although interestingly this happens less often compared to i2i. I'm not sure what has caused it. I've tried downgrading ComfyUI, fresh installation of ComfyUI container, all sorts. No luck. This randomly started happening while generating images yesterday morning, suddenly this noise. With t2i workflow I'm at least able to get a proper image output every 1 in 4 generations or so, the rest have noise or glitches.
If you find the solution I'm all ears, I'll also post if I find a solution, but so far I'm stumped. I also asked on the ComfyUI Discord server, one person replied but it wasn't a solution, so no help there. Given it's not likely my GPU is broken (RTX 5090), at least some simple math tests passed with expected results, At the moment I can only assume that it's either something in ComfyUI or something that ComfyUI is using. I've even tried downgrading torch.
Make sure you use the sampler of euler-a/beta recommended by Phr00t, other samplers might have side effects. Install any attentions as possible as you can, even you have 32GB Vram or more. Some loras might cause the issue too if you use them.
Make sure you use the sampler of euler-a/beta recommended by Phr00t, other samplers might have side effects. Install any attentions as possible as you can, even you have 32GB Vram or more. Some loras might cause the issue too if you use them.
Hi, many thanks, but in my case it was definitely an unstable core clock speed (at default). Even on ComfyUI default workflows for other models, the templates, I was getting similar issues. I can only assume my card has had some kind of minor degradation. I've done further fine tuning and the issue remains resolved if I reduce core clock offset by 50 MHz, default memory clock is fine, and power limit can also be 100% although I prefer to keep it around 75% as I don't notice any real difference in performance when compared to 100% power limit. I'll keep an eye on it though, if it re-occurs and I have to reduce the core clock speed further then I'll have to consider potentially RMA'ing it assuming it's still in warranty by then.