I’m not posting as much as I would like to, been fighting cancer for a year now. Reallusion products are one of the reasons keeping me active and creative during these dark times. I just wanted to share what I’m thinking of AI Studio, the good and the less good. Helps take my mind off things
So I’ve been playing around with some of the digital doubles I made in the past months. Reallusion didn’t lie regarding image and still generation: it’s overall impressive and pretty fast (and doesn’t consume too much of GPU power). Personnally I’m more drawn to photorealistic rendering, so this post will focus more on that for now.
What I’ve seen is that handles really well front face rendering, giving a nice photorealistic look. Regarding facial animation, I could run a short clip (234 frames) and it didn’t disappoint either. Everything was there: eye movement, wrinkles, realistic hair… All extremely promising.
As AI is upgrading at lighting speed, I’m eager to see how it will be upgraded in the future.
Now for the less good… it’s gonna be the same as always: pricing. EVERYTHING comes with a price. I understand that an image generation costs points. But enhancing the quality, 40 credits for each picture ? Seems a lot. Why don’t start with a 2K generation instead of 1K ? Anyways, I’m complaining but I know that’s how every AI tool works.
Also, generating AI images from 3D renders starts to get tedious by the third or fourth attempt (especially the right-side view, which almost always ends up as a left-side view—I don’t know why).
Face generation is impressive and very believable. But patterns or textures of 3d clothes are almost never preserved once the AI-generated version is complete. Which can cause consistency issues.
And that’s the last point: consistency. It is almost and always the same issue with AI. 10 seconds max of video generation, but the results are at their best if we keep 2 or 3 seconds. For now it’s very difficult to work for an animated movie with such a short time of trusted results. If too long, faces can distort in a funny way, lights can get oversaturated or completely change from one frame to another.
I think it’s really promising as for now (and I am a complete anti-AI artist, which is why I was very cautious) but I’m especially looking forward to the new RTX rendering engine.
If AI products like Seedance continue to evolve and Reallusion keeps up with the pace, I think the combination of these two tools—AI and the rendering engine—will be able to preserve artists’ work while giving them total creative freedom and adding visual effects worthy of a major Hollywood studio.
But for now, we’re not there yet.
I’m posting some of the renders I’ve been playing with this tool (3d models of Sofia Boutella, Henry Cavill and Navid Neghaban). I’m planning on keeping Iclone native render for now.
Over the last couple of years all big online AI services were subsidized to grow a user base and the companies behind them lost billions of dollars each year. They now need to generate income or otherwise the investors will lose a lot of money.
This means that tasks which need a lot of compute like AI agents and AI video generations will become increasingly expensive or will reduce the feature set because providers will switch from flat subscription fees to usage based fees. There are enough reports online where Github Copilot power users, who previously paid $200 a month, now would need to pay $2000 or even $5000. Other users run into newly introduced and very strict usage limits.
Sadly, it looks like AI Studio currently doesn’t support local models which could save a lot of costs, depending on whether you already own capable hardware.
As far as I know, a lot of AI video generation models are trained on a limited frame number, for example 81 or 121 frames (which are stored as an 9x9 or 11x11 image grid in VRAM during generation). Increasing the number of frames naturally degrades the video. You can technically stretch it longer with a slow motion effect in your video editing software or by concatenating multiple videos.
It is certainly understandable why Reallusion doesn’t support local models in their AI hybrid systems.
All of the local models are woefully underpowered and produce really subpar results even with still images and they are not even close to being competitive with video generation or anything similar to an agent base system.
@AutoDidact To me it looks like AI Studio supports LTX 2 and 2.3 and Wan 2.2 Fun Control, which are available for local usage. But it seems here they are running on the server backend.
I would guess that RL is renting their compute at a volume discount and upcharging the Iclone/AI users to make a small profit
via a token or credit pay per use system
thus would have little incentive to encourage people running local compute.
I have not been playing with AI Studio but I have used AI for the last half year or so using a service I’m subscribed to. For character consistency it is essential to create a series of images with the character from different angles (like a LoRA). My service uses 10 images for that purpose. Doing that really helps. That should have to be something for AI Studio to support to make it useful.
I usually start with a rendered image in the style that I want (which is semi-realistic) as start frame and then possibly a second one as end frame. So it sort of a hybrid model where I use iClone and CC to create the scene, render it with Nano Banana 2 and then use Kling 3.0 to create the video. The service also provides a few free models, which is good for tests. I mostly render in 720p and the scale up, which works quite good overall.
As with everything, it takes time and effort to arrive at what you want.
I’m on travel right now so I’m not in a position to evaluate AI Studio.
I don’ know why it was released now, before RTX Render, so I’m not quite sure where RL is heading with this.
That works for creating the AI actor sheet, but from my test it doesn’t seem to carry through properly into 3D-to-video. I can create the actor and assign it in 3D-to-image, but the video result still doesn’t preserve the actor consistently. I also lost credits when the first attempt crashed, and on the second attempt the photorealistic style only appeared for about two seconds before the rest looked like my original 3D video again, with strange generated sound effects. I lost 300 credits doing that. So the actor sheet is useful, but it doesn’t solve the 3D-to-video consistency problem yet.
We love seeing the discussion here. To make sure we see all your ideas and help troubleshoot issues faster, you can share your thoughts, creations, and questions over in the official AI Studio Category
Check out our FAQ resources in Official and find out how you can earn free AI Points just by sharing your work in Showcase & Earn!
Update: This thread has been moved to the AI Studio Category
To be honest, it’s crazy expensive. They need to figure out a way to reduce the costs involved. With all the errors and Re-trying when using AI, a minute video would be very expensive.
Update: This thread has been moved to the AI Studio Category