Yes, switching to OFX as well, way more potential. The inference node current implementation is 100% banned for us.
[edit]
And this is key:
“I think it’s about official frameworks that allow to integrate those API keys and harnesses to control them in ap alongside pipeline-esque modalities for making the results feel like an integrated piece of the kit.”
Is that for AI tools to do the whole shot, or for AI to be part of the pipeline and parts of a shot?
Everything I’ve seen of ‘prompt for the shot’ has indeed fallen short, unless you are super flexible about the story, framing, etc. which is the opposite of what our jobs are. And it seems unlikely to change dramatically.
But there have been good examples of AI being part of a timeline and saving time, or unlocking something that may otherwise not happen.
Yep. It’s not super difficult either, just takes time…
Whole shot tools are none of my business, and we can’t use them anyway, so I’m not talking about those.
Topaz, Silhouette/Syntheyes new roto tools, and the new time warp tool help me occasionally, but those are ML, not AI (though to be fair, not of it is ACTUALLY AI, but that’s for another thread). Are they doing things that weren’t otherwise possible? I would not characterize these apps that way, they are occasionally convenient and just as likely to be frustrating.
True final pixel roto remains elusive and the 85% or less that the existing tools leave me with is often no better than just doing it the old fashioned way. It’s rarely even good enough for color.
ML timewarps are great, but almost always need a significant amount of paint work to get to final, which isn’t any fun.
Gaussian Splats are interesting, but there aren’t a lot of great tools out there and cleanup of the textures is a pain, so it remains a hobby for me.
Things like infill or cleanup continue to be tested, deemed promising, and then discarded when they fail on fairly straightforward tasks. I put nuke copycat in that category. Cool demo, not terribly useful for our purposes.
Other folks in other parts of the organization love the ability to iterate quickly and try out ideas without having to build or shoot or code anything, and I love that for them, but I deliver finished work, any frame of which could be paused on and snagged and used every time a post or news story mentions the CEO (purely hypothetical example), so my patience for “pretty good! Imagine how great it will be in the future!” is down to about zero.
I would love for these tools to work, and to help me, I have no ideological axe to grind against ML/AI that assists artists. But most of the development focus seems to be aimed at photos and full generative scenes, and those don’t really help me.
Makes sense. Everyone’s mileage varies a bit depending on expectations. Your’s are at the upper end of the scale for good reason.
Agreed with a lot of what you said, but I need to give props to the automatte in 2027. We used it extensively on every shot for multiple spots, and it was great.
This was for grading/look dev — not comping though. Some took a little manual fixing, but the time and cost comparison to going to traditional roto — this won this time.
I think the issue with outpainting (esp video) is that you are running up against the limitations of the tech. You can only train an image or video gen model on so many examples, and there’s literally infinite combinations of what is possible. I can’t believe I’m going so say this, but there’s only so much content on YouTube to scrape and train 
. Then there’s the resolution and bit depth constraints, and the cost of compute. I think this is why the roto and roto refine models, depth, etc, really show the limit of what is possible with a general tool. I’ve gotten much, much better results from locally trained models on very limited data sets, but those are pretty much bespoke to the given project.
I’d def like to see a way to hook more models, etc, into flame, but I want the devs at ADSK to use their resources on things that will work, reliably, every time. 
Batch Paint entered the chat.
OK, how about if I reword it to: ML tools?
Did your prompt for this specify “thigharrhea” or did you inadvertently prove my point? 
I’m guessing that automatte saved me 2-3 hours of roto yesterday. I use it almost every day.
Interesting thread. Also find it intriguing that people’s use case and success using AI largely differs.
we use Flame’s ML timewarp often and in the majority of cases haven’t hadn’t had to do much or any cleanup of it. Before someone suggests we’re not looking hard enough, we’re delivering work to the major streamers and studios so QC is about as detailed as it gets and we are not getting kickbacks for any ML timewarp issues.
We’ve been using AI/ML roto for certain types of shots and it has worked intermittently. When it does work it is a timesaver. We actually used 2025’s ML face detect for some beauty work that there was very limited budget for and it worked a treat. Have done similar on beauty work previously too and the work wouldn’t have happened at all if it weren’t for ML so hasn’t done anyone out of a job but has earned some extra revenue and even lead to more work for us.
We have used InPaint and CopyCat regularly and when it works, which is in the majority of cases, it has saved time and money.
We’ve used some AI generated depth mattes for a few shots that has led to some excellent results.
I’m using AI to help develop and improve our pipeline. We’re not a big enough team to be able to afford a pipeline TD and I don’t have the time or skills to code it myself. We’re getting some real efficiencies building custom tools for the way we like to work, lean and mean.
Are there cases where AI doesn’t work. Yep, shitloads (technical term for lots). But we initially give it a burl to see if it will get us somewhere and if it isn’t in the ballpark we quickly drop it and switch methods. Yes, it can produce work that isn’t up to standard but that’s where “human-in-the-loop” thinking steps in. You reject bad work regardless of how it is made. I’m definitely not an AI evangelist, I’m just the type of person who will pick up a tool that I think is right for the job. Sometimes I change approach and pick up a different tool and I’m thinking about some recent DIY plumbing I’ve had to do, not working on shots
I just wanted to post this because just because one person’s experience is underwhelming doesn’t mean that it isn’t working for others nor that their idea ls are flawed. I’ve been seeing that a bit on Logik as of late which is quite sad. People’s use cases and experiences with Flame are pretty diverse which says a lot about the versatility of the software and what the dev and support team have to look after with finite resources. I just want to say thanks to @Slabrie @fredwarren and the team for their upkeep of what must be a real beast of software to maintain and upgrade. Must feel like working on a classic car at times.
@finnjaeger You may have already seen this but is this the sort of thing you were suggesting for Flame? On the surface it looks pretty interesting
Can I also suggest that maybe it is the likes of Highsfield we need to be asking to build tools instead of a limited dev team to build for Flame?
My concern for Flame though is, and I’ve expressed it before, if you start utilising external tools that have their own costs then people will start looking to the likes of Resolve to lower their operating costs.
I will say that implementing this via a socket and cli is trivial if you’re working unmanaged. Since Higgs runs with Oauth you can use a personal account and not fuck about with api keys.
I’m doing something similar for a project I’ve been working on. If there’s interest I can tear it out as a sep repo.
It’s contributions like this that make Logik rock.
At this stage we are not looking to use Higgsfield as there are too many restrictions on the use of GenAI for us as a business but I bet you there’s lots of folks on here that would love you to do that.
Done 
Not exactly native code from the Flame devs, but gives a production ready toolset for the kind of helpful AI. Indeed it ships with Inpaint, Outpaint, and SAM nodes. I’m due to make another video about the shioping contents, but got hella busy. There is an older one or two vids about it on my youtube channel ( is 3 vids make a channel).
“https://www.youtube.com/@Beakfx”
they are from when I was still building it, but still relevant. Just more polish and more nodes now.
-----Beak
Autodesk lovet to buy apps foundry just bought griptape to help bridget the gap in AI. Autodesk can find a way and do the same.
i’ve been exploring AI tools for enhancing my workflow too and found that creating a tailored solution can be more effective than off-the-shelf options. for those considering this, redeagle.tech offers resources that emphasize the importance of aligning AI solution with actual user needs during development
this might help mitigate common frustrations liiike AI tool limitations by ensuring the solution effectively addresses specific visual effects challenges