other tests
Hi.
I have made other tests to get better character consistency. I tried to push the strength of each picture in LTXVAddGuideMulti around 1.5. No change.
In the fourth picture I loaded a dog. In the render it is completely different. No shape, color, or texture is consistent with the reference picture. Another quite strange point: I noticed that in your prompts there is no reference to the original picture number. Generally in multi-reference workflows with Flux, for instance, it is required to make the relation between a character in the prompt and its picture number.
Third, I noticed that your reference pictures are all high definition and have the same definition and aspect ratio. Is it important and mandatory?
Another question: can I use model sheets with multiple views of the same character in one picture?
Thank you very much for your help. I hope you will publish soon a definitive version of this great tool.
According to my understanding of LTX2.3 code, the strength shouldn't be set greater than 1.
Using multiple reference images of the same character requires addressing the problem that the model does not treat them as different subjects.
Seedance 2.0:
Hi.
I have made other tests to get better character consistency. I tried to push the strength of each picture in LTXVAddGuideMulti around 1.5. No change.
In the fourth picture I loaded a dog. In the render it is completely different. No shape, color, or texture is consistent with the reference picture. Another quite strange point: I noticed that in your prompts there is no reference to the original picture number. Generally in multi-reference workflows with Flux, for instance, it is required to make the relation between a character in the prompt and its picture number.
Third, I noticed that your reference pictures are all high definition and have the same definition and aspect ratio. Is it important and mandatory?
Another question: can I use model sheets with multiple views of the same character in one picture?
Thank you very much for your help. I hope you will publish soon a definitive version of this great tool.
Thank you for your feedback.
Yes, my LoRA does not currently establish an explicit mapping between the image number and the prompt. So for now, you need to describe the visual features of the reference images as clearly and specifically as possible in the prompt to help maintain consistency.
In the next step, I may try to develop a numbered-reference mapping approach.
Also, the aspect ratio of the reference images must match the target video. In fact, my node automatically resizes the input images to exactly the same resolution as the target video.
Third, yes, you can use a multi-view model sheet of the same character in one image.
Perfect. Please update your README file to mention the resolution and aspect ratio, which are mandatory. And to describe the objects and characters in the prompts because normally we must AVOID describing them because it can confuse the model. It will help people to avoid bad results.
Here's a snapshot of a test I made of your workflow just cropping your cat with less resolution. It generates two cats and does not follow the prompt.
Perfect. Please update your README file to mention the resolution and aspect ratio, which are mandatory. And to describe the objects and characters in the prompts because normally we must AVOID describing them because it can confuse the model. It will help people to avoid bad results.
Here's a snapshot of a test I made of your workflow just cropping your cat with less resolution. It generates two cats and does not follow the prompt.
Thank you, since the v1 version is coming soon, and some issues has been significantly refined.So, I'll update the model card with the new version together.
LiconStudio when you can update new version to fix the plastic look? Thanks you so much
LiconStudio when you can update new version to fix the plastic look? Thanks you so much
Not sure, I'm still testing different tragic, but it wont be long since I'm testing on 40 H200s.
Thanks a lot for you work! Can you share how many GPUs do you use this training. 40 H200?
Thanks a lot for you work! Can you share how many GPUs do you use this training. 40 H200?
40 H200s, but most of them are for several different tragics' testing, thats why V1 comes up in only a week.