dx8152

/

Qwen-Edit-2509-Multiple-angles

like 770

This is truly excellent work—thank you for sharing it!
May I ask whether the training code you used is identical to the open-source version on ModelScope, or did you make any modifications?
Also, could you tell me which dataset you trained on? Thanks again for your generous contribution!

I'll be releasing a video soon explaining the training methods. The model is the same as ModelScope; LoRa is trained using ModelScope.

Did you create the dataset over Unreal Engine? @dx8152

Did you create the dataset over Unreal Engine? @dx8152

No, I used real photos, which made it difficult to control some angles because it was hard to find the corresponding photos.

Does something like this require a lot of photos for dataset? Ex: few hundred for each of the angles you want.

Also for getting the dataset and finding it hard - you could breakdown videos as that would a) get you the angle b) give you scene consistency

Does something like this require a lot of photos for dataset? Ex: few hundred for each of the angles you want.

Also for getting the dataset and finding it hard - you could breakdown videos as that would a) get you the angle b) give you scene consistency

Video is indeed a great method, but it wasn't considered during production.

Regarding the quantity, there won't be that many. For example, one or two pairs of photos rotated 90 degrees to the right would be sufficient.

Does something like this require a lot of photos for dataset? Ex: few hundred for each of the angles you want.

Also for getting the dataset and finding it hard - you could breakdown videos as that would a) get you the angle b) give you scene consistency

Video is indeed a great method, but it wasn't considered during production.

Regarding the quantity, there won't be that many. For example, one or two pairs of photos rotated 90 degrees to the right would be sufficient.

@dx8152 , so you trained this lora with around 20-30 photos only? 🤔

Does something like this require a lot of photos for dataset? Ex: few hundred for each of the angles you want.

Also for getting the dataset and finding it hard - you could breakdown videos as that would a) get you the angle b) give you scene consistency

Video is indeed a great method, but it wasn't considered during production.

Regarding the quantity, there won't be that many. For example, one or two pairs of photos rotated 90 degrees to the right would be sufficient.

@dx8152 , so you trained this lora with around 20-30 photos only? 🤔

We need around 50 pairs, but this data is quite difficult to find; we only have about 40 pairs of images available.