[Official] 🎉 Welcome to AI Render Open Beta – Start Here!

Open Beta AI render
126

I want to start by saying THANK YOU REALLUSION FOR LISTENING TO YOUR CUSTOMERS!!! I am really enjoying the new render engine. I only have 2 points of feedback:

  1. The render doesn’t resemble the model as much as I would like, I guess that is probably a learning curve on my end, but just voicing feedback.
  2. is the fact that it takes 15 minutes for a normal render on my computer, and about an hour for the photorealistic render, granted my laptop is from 2019, even though it still has plenty power and memory

    Result03
    Result031920×1080 48.8 KB

    RenderImage3
    RenderImage31920×1080 38 KB

    Result01
    Result011920×1080 76.3 KB

    RenderImage1
    RenderImage11920×1080 128 KB

    Result02
    Result021920×1080 72.9 KB

    RenderImage2
    RenderImage21920×1080 82 KB
127

Try experimenting with the guidance value (CFG).

Has your laptop an RTX GPU with enough VRAM? If the rendering process doesn’t fit in VRAM it is processed on the CPU which runs at least a magnitude slower.

128

If the laptop is from 2019 then that would be pre-RTX, I believe. I would be interested in the specs because while on vacation I only have an older laptop (Alienware 17 R5) that I would like to try it with.

129

RTX came out in 2018. But it may work as well with a GTX card. I have an older GTX 1660 Ti which can run Stable Diffusion as well.

If you’re on vacation you can also try to access your main computer via Parsec. This is a remote access app that is optimized for gaming or 3d applications. But then you need to keep your PC at home running all the time. It might technically also work via Steams Remote Play, but it has certain security measures that might block working with desktop apps.

130

Thank you for the tip. My PC (with an RTX 4090) is off right now but this would certainly be worth looking into. I have a good internet connection here.

131
Device name DESKTOP-08SBLDB
Processor Intel(R) Coreℱ i7-10750H CPU @ 2.60GHz (2.59 GHz)
Installed RAM 16.0 GB (15.8 GB usable)
Device ID 586064AC-939C-4111-B8A4-A8EBB244131D
Product ID 00325-82055-15602-AAOEM
System type 64-bit operating system, x64-based processor
Pen and touch No pen or touch input is available for this display
132

OK, so as far as I figured the whole AI render is wrapped around face detection.
If it did not detect a face, it does not render. I mean it does, but it comes out jagged, blurred, distorted, smudged or all of the above.

iClone render of a basic scene:

1
11280×720 44 KB

Same rendered with AI. Settings do not matter at all, it would come up pretty much as this:

2
21280×720 238 KB

Now I added a tiny face of Camila at the corner and everything falls in place:

3
31280×720 252 KB

I hope it fixed at some point


1 Like
133

Here is how it all started

I was trying to render some scene for hours and it kept coming out like this:

4
41920×1080 146 KB

Then, when I finally realized what needs to be done, I got help from a base female bystander :slight_smile: She could be easily “clone brushed” out though.

5
51920×1080 187 KB

1 Like
134

@Sp3c1alk: What is your graphics card?

137

Hi 4u2ges,

This issue is most likely caused by the Consist Expression feature in Live Portrait (located in the Control section) failing to properly detect the face.
If your generated output isn’t meant to focus on facial expressions and you’re seeing distortion in the image, try turning off this option

image
image594×665 21 KB

1 Like
139

Hi guys,

I did try that before and it did not work either.
Here is a simple test:

140

Hi there, I’d suggest trying a resolution of 1024x1024 to see if it speeds things up.
Since our model is based on SD1.5, using a square resolution can help improve processing efficiency.

Ask ChatGPT

141

That’s an interesting find!

142

Thank you for your feedback. We’ve recorded the issue and will include a fix in the upcoming update.

143

We appreciate your concern.

Your post was hidden due to community reports and forum protocol. We’re sorry if this caused any offense.

It’s difficult to fully control results from open-source models, and some prompts can indeed lead to inappropriate outputs. We include default negative prompts to help reduce this risk, but they aren’t 100% effective.

Ultimately, it’s up to users to review and take responsibility for the content they generate and choose to share — this isn’t an issue unique to AI or 3D tools alone. Thanks again for bringing it to our attention🙏

1 Like
144

How can a user generate responsibly when the software is spitting out this content with default settings? If this issue is “Known” to AI
 why were there no safe guards built in?

I’m sorry but to take an adult toon character and turn it into a nude baby should not happen. A nude adult toon is fine but not a baby or child.

I mean all I did was to select the style and hit preview. I should not get such results from a default setting.

I am sorry that you found the results of this software so offensive in its covered state. So much you removed it. Imagine how I felt when it generated it in front of my eyes.

It’s ok, you have your rules however I find it funny that other sites have not found such offense. They can see what dangers can await them with AI.

It really is a shame you do not care enough about your users to give them the same warning. :frowning:

145

Jason_RL noted your concern, but also pointed out that RL added safeguards already as far as possible. However, the AI model, or any other one, is not under RL’s control, so they are not able to change the model directly.

There is an art in coming up with a prompt that cannot be misinterpreted, I have found that out myself, and then the results can be unexpected. This is a very new technology that is not well understood.

I also think it goes a bit far to claim that RL does not care about its users based on one incident.

1 Like
146

Let’s walk through what you actually did — step by step — since you seem a bit confused about how Stable Diffusion 1.5 works in open-source environments like iClone/ComfyUI.

1. You used SD1.5 — a small base model that is lightweight to ship with no built-in NSFW filter. This isn’t some cloud platform with massive regulation watching your every move while harvesting your intellectual property. SD1.5 is an open-source model trained on massive amounts of scraped internet data. That includes NSFW content, adult themes, and everything in between. It does not stop you from generating something unethical — because it was never designed to. YOU, the user, are expected to control that.

2. You supplied a prompt — improperly.
From your post and the results, it’s clear you put your positive prompt into the negative field. That means instead of saying “avoid adult content,” you told the model: “Avoid everything except adult content.” You inverted the mask. That’s not a bug — that’s user failure.

3. You ran the model with those instructions and got disturbing output.
And instead of owning that mistake, you blamed the interface.

4. The result was not default behavior — it was directly caused by your input.
No model, not even SD1.5, “accidentally” makes a toon character look like a sexualized child. That requires specific prompt elements that came from you. There is no “default style” that leads to that. You either prompted it, or your mishandled prompt did.

**5. You posted it — knowingly **
You uploaded it to a public forum. That’s not a preview accident. That’s intent. And the fact that you then turned around and tried to frame it as a lack of safeguards shows exactly how manipulative this post is.

Bottom line:
You used a powerful tool irresponsibly. You failed to understand basic prompt structure. You generated unethical content and then blamed the tool when it did exactly what you told it to do. And when someone moderated it, you lashed out as if you were wronged.

If you’re not willing to take full ownership of what YOU typed into an open-source model, you should not be using that model. This isn’t “just click preview.” This is an industrial-grade AI system with a easy to use but professional cg front end pipeline. Treat it accordingly please, or hold off a until the official tutorials launch to help you understand how to prevent such things from occurring due to your current lack of understanding of how AI models work and their specific capabilities.

Soon there will be a release of intro tutorials and webinars for you to install something like SD3.5/Flux or other models that have built in moderation while you overcome the basics of using features like prompting, ControlNet, and LoRA training.

2 Likes
147

“Improved” version is ok if you dont mind changes but if I want “my character” will the lora keep it intact?

148

Hi ender570,

For those who wish to use their own ComfyUI environment, please refer to this reply as an initial reference. A more comprehensive document will be available soon.