Hey guys, this might be too broad a question, but how does Chaos group view traditional offline rendering going forward? Is the market expected to shrink drastically, or is it a case of adapting things like Comfy UI style workflow into the Vray pipeline?
Ive seen some pretty impressive stuff with basic cg scene outputs like normal maps and depth maps being piped into a Comfy workflow and generating content using these utility passes. Results that maybe you have less control over, but are there in a matter of seconds, rather than the time it takes to create the shot (hours / days). I feel like there is probably a hybrid middle ground there where maybe raytracing rendering is a thing of the past, and having a chaos trained LLM create the final output is maybe where its headed? Any thoughts?
Hi!
I am bumping this. What do people and Chaos think about this? I guess there must be a lot of strategical decisions taking place in Karlsruhe and Sofia these days.
Yeh this didnt gain much traction, but maybe its wise to play their cards close to their chest. Volatile times in the industry and theres probably a lot to risk by making claims one way or another. Some people see AI as the new normal, some see it as a moral and ethical issue…both have validity.
One thing that isnt related to Chaos’s plans but they might be able to give some insight on, is whether they know if AI image gen and AI video gen is a loss making enterprise? Ive read a bit lately that all of the tools we are being told to adopt are heavily subsidized currently, and cant continue as it stands. So either the tools reduce in quality, or the price goes up…ideally for them once the industry has abandoned its current tools and is reliant on the AI workflows.
Basically this amounts to the Uber model…lose money until market adoption is huge and then start the enshitification process. Has Chaos Group looked into any AI model options and seen the numbers one way or another? Id love to know if these tools are a flash in the pan or not for image creation. Obviously the AI models could improve in efficiency and quality, but I also read that in their current form, its basically a matter of scale…to make bigger and better models, in the format they are now, just takes more compute power, and there is a limit to that in terms of cost vs return.
we might be in the point in history where modernizing/rethinking energy infrastructure is falling way behind progress in AI. we need more compute, compute needs more power and power is not getting cheaper fast enough. recent events have also shown some of this compute will need to happen closer, in high-energy-cost places like Europe. I think it’s just weird times right now and some correction is about to happen.
also I would not bet on Chaos sharing bussines strategies here - but nice to have a discussion nonetheless.
there are people claiming that the future of AI is not cloud but local - not sure how that changes the landscape but price of 24+GB vram gpu seems now astronomical compared to “cents per generated image” pricing model. cloud compute is just like photoshop going subscription route. progress I guess.
I’ve done some production ready AI visuals recently (archViz). Each started with a basic flat render from Corona, essentially a diffuse render element! Running Nano Banana Pro 3, comfy UI.
Accuracy was an issue, small details mainly but you get around it. Ever played whack a mole?
Updating the visual with a revised model render was tedious but possible
I spent a lot more time in Photoshop than I usually would (quite refreshing actually)
For vegetation, like shrubbery you can just use blobby meshes in the base render and get the AI to generate realistic replacements (way better than using chugging chaos/FP scatters)
Final images probably took the same amount of time to complete but you reach a higher quality much faster and you can take the image much further.
Clients also get a much better idea of where the image is going sooner.
Going back to a frame buffer seems archaic.
I think when image AI’s stop downscaling the image you give it and thus reduce the data it has to work with (less pixels, less details), I believe a lot of the concerns over quality and accuracy will dissapear.
all the publicly released models you can run locally provide a good baseline for what should always be possible from now on - the only resource you need is decent gpu.
I havnt checked in on the open source models for a few months, but I didnt find anything came close to Nano Banana pro 4k for image edit, Seedream 4k for image gen. Everything I tested was natively sub 2k…and I didnt even test video cos on my 3090 it just takes too long for 1080p. Do open source models now handle text and logos well? Have they moved beyond 1024px being the norm?
Because of the inherent one armed bandit, slot machine style interaction with image gen models, I often need to run off a dozen images to get the right result, and on one machine, with one GPU, needing 4k+, then needing to upscale to 10k, that just doesnt compute time wise on my local machine when client needs to see something in an hour.
few months is ages in this environment but obviously open models are well behind the cutting edge. this yt channel has very decent info about technicalities of ComfyUI: pixaroma - YouTube - how to get the most possible control over generations/editing (open source models or not - Comfy makes it a lot easier).
on 3090 - zimage turbo bf16 with 9 steps took a bit less than 1 minute to generate 1664x2432 pixel image. most likely not possible to generate 4K I guess but speed is acceptable.
Yeh, I did dabble with comfyUI for a month or so at the start of the year. Ended up using NanoB pro with API key in Comfy since nothing else could edit like it can…but there is a low file size limit for image edit uploads…nothing over 5k or something. They said theyd fix it not sure if they have yet or not, but there is no limit (that ive found yet) in apps like Google AI studio or even Freepik, so I switched to them.
I like how some open source models let seem to let you plug in more reference inputs for video, like a basic render, depth, and edge passes. When Ive used paid for models online, Ive only ever been able to use one mp4 file to drive the AI output…feels limiting.
Thats super interesting, will defo check them out. I was pointed to Art Craft by someone yesterday, also looks interesting. Seems to be made by a VFX / motion graphics guy and geared towards solving some of the gap between AI and VFX. First thing Ive seen that lets you layer up assets and use them in a 3d space, then run them through image / video gen.
Most offerings so far (higgsfield / freepik) seem to be catering to web content creators and social media types.
