3 ms·

I think you're discounting the possibility that this can be (at least partially) automated. Specifically things like changing sponsor logos or animation models

I think you're discounting the possibility that this can be (at least partially) automated. Specifically things like changing sponsor logos or animation models. I can envision tools that allow a sign in a store front to change from "Coke" to "Pepsi" automatically, or even replace that Audi car the hero is driving to a Mercedes. Of course, not everything could be automated but I think a good portion of it could be. The original VFX team would just have to insert "markers" into the workflow which could then be replaced and re-rendered. Almost like changing build flags and recompiling a program.

the8472 7y ago
In fact this already already happens, currently it's used the ads industry for luxury car segment where they render individual ad reels for any car configuration you order.
TeMPOraL 7y ago
The other day someone sent me this: https://www.youtube.com/watch?v=jLg5CZ-AG0g&feature=emb_title https://www.youtube.com/watch?v=jLg5CZ-AG0g&feature=emb_titl... This shows a product that apparently can scan a video for good places to put ads, and then replace or insert advertisements into it. And by "insert" I mean "add branded cups on the table where none existed" or "add a billboard to the cityscape where none existed". Scary stuff, I absolutely hate that people are working on such things.
mroche 7y ago
> I can envision tools that allow a sign in a store front to change from "Coke" to "Pepsi" automatically, or even replace that Audi car the hero is driving to a Mercedes. Product visualization/ads has had pipelines set up for this near drag and drop for static assets for a while now. Their assets are known before production, and have been decided by the agency that made them. Animation and Film VFX are setup fairly differently. The ability to make changes like that already exist. Scenes and assets built procedurally can be altered with relative ease as long as they don’t encroach on other aspects of the shot. > Specifically things like changing sponsor logos or animation models. Logos and for the most part basic comp-only changes? Yeah that’s fine. Animation model is a completely different story. Changing the base model means the UV’s likely changed so the texture artists will have to either update (can be complex if they’re trying to maintain their original stack) or redo their work. It means the rig needs to be updated and modified. If it’s a muscle based system for precise and realistic skin deformation, have fun. If there character has fur, then the grooming team has to do another take on the asset as well. Animation then needs to make another pass on their scenes to make sure the animation actually looks good and lookdev needs to make tweaks to maintain the look they had before. And then you have to rerender, which will likely conflict with the rendering allocations (if using an on-prem farm) of the current deliverable unless the studio is in a lull point. And rendering takes time. Then comp gets to deal with the change and make sure it still sits in the shot. And we still haven’t discussed the time and tweaks to re-sim and cache any affected effects. For a movie like Cats, this is going to be a lot of work, with a ton of anim, matchmove, and comp fixes. > The original VFX team would just have to insert "markers" into the workflow which could then be replaced and re-rendered. Almost like changing build flags and recompiling a program. That is a significant misrepresentation/and lack of understanding of how these pipelines work and function. No matter how much you can automate, this is not just data the way the typical HN user sees things. This is someone’s art and vision, and will always require a hands on approach with human eyes. If it was as simple as changing a few parameters or altering a few lines of code, there would be far fewer people in the industry.
krut-patel 7y ago
> I can envision tools that allow a sign in a store front to change from "Coke" to "Pepsi" automatically. This was literally the project I worked on this summer, as an application of GANs for "image translation". We basically trained some of the latest GAN models on our dataset of Pepsi and Coke bottles. Thankfully, all the research papers are accompanied by complete tensorflow/pytorch implementations so we only had to do minor modifications to train/test them on our dataset. The results were mixed, it worked pretty well for bottles on white backgrounds, and one network properly learnt the importance of orientation of the logo (horizontally flipped images generated corresponding flipped target logos). The field is pretty nascent, and from the papers I've read, NVIDIA Research seems to be making good progress on this. I believe in two or three years, we can definitely see this tech in production.

the8472 7y ago

In fact this already already happens, currently it's used the ads industry for luxury car segment where they render individual ad reels for any car configuration you order.

TeMPOraL 7y ago

The other day someone sent me this: https://www.youtube.com/watch?v=jLg5CZ-AG0g&feature=emb_title https://www.youtube.com/watch?v=jLg5CZ-AG0g&feature=emb_titl... This shows a product that apparently can scan a video for good places to put ads, and then replace or insert advertisements into it. And by "insert" I mean "add branded cups on the table where none existed" or "add a billboard to the cityscape where none existed". Scary stuff, I absolutely hate that people are working on such things.

mroche 7y ago

> I can envision tools that allow a sign in a store front to change from "Coke" to "Pepsi" automatically, or even replace that Audi car the hero is driving to a Mercedes. Product visualization/ads has had pipelines set up for this near drag and drop for static assets for a while now. Their assets are known before production, and have been decided by the agency that made them. Animation and Film VFX are setup fairly differently. The ability to make changes like that already exist. Scenes and assets built procedurally can be altered with relative ease as long as they don’t encroach on other aspects of the shot. > Specifically things like changing sponsor logos or animation models. Logos and for the most part basic comp-only changes? Yeah that’s fine. Animation model is a completely different story. Changing the base model means the UV’s likely changed so the texture artists will have to either update (can be complex if they’re trying to maintain their original stack) or redo their work. It means the rig needs to be updated and modified. If it’s a muscle based system for precise and realistic skin deformation, have fun. If there character has fur, then the grooming team has to do another take on the asset as well. Animation then needs to make another pass on their scenes to make sure the animation actually looks good and lookdev needs to make tweaks to maintain the look they had before. And then you have to rerender, which will likely conflict with the rendering allocations (if using an on-prem farm) of the current deliverable unless the studio is in a lull point. And rendering takes time. Then comp gets to deal with the change and make sure it still sits in the shot. And we still haven’t discussed the time and tweaks to re-sim and cache any affected effects. For a movie like Cats, this is going to be a lot of work, with a ton of anim, matchmove, and comp fixes. > The original VFX team would just have to insert "markers" into the workflow which could then be replaced and re-rendered. Almost like changing build flags and recompiling a program. That is a significant misrepresentation/and lack of understanding of how these pipelines work and function. No matter how much you can automate, this is not just data the way the typical HN user sees things. This is someone’s art and vision, and will always require a hands on approach with human eyes. If it was as simple as changing a few parameters or altering a few lines of code, there would be far fewer people in the industry.

krut-patel 7y ago

> I can envision tools that allow a sign in a store front to change from "Coke" to "Pepsi" automatically. This was literally the project I worked on this summer, as an application of GANs for "image translation". We basically trained some of the latest GAN models on our dataset of Pepsi and Coke bottles. Thankfully, all the research papers are accompanied by complete tensorflow/pytorch implementations so we only had to do minor modifications to train/test them on our dataset. The results were mixed, it worked pretty well for bottles on white backgrounds, and one network properly learnt the importance of orientation of the logo (horizontally flipped images generated corresponding flipped target logos). The field is pretty nascent, and from the papers I've read, NVIDIA Research seems to be making good progress on this. I believe in two or three years, we can definitely see this tech in production.