Why cant we use 2 GPU's the same way RAM offloading works?

I am in the process of building a PC and was going through the sub to understand about RAM offloading. Then I wondered, if we are using RAM offloading, why is it that we can't used GPU offloading or something like that?

I see everyone saying 2 GPU's at same time is only useful in generating two separate images at same time, but I am also seeing comments about RAM offloading to help load large models. Why would one help in sharing and other won't?

I might be completely oblivious to some point and I would like to learn more on this.

https://redd.it/1l6j4y9
@rStableDiffusion

Check this Flux model.

That's it — this is the original:
https://civitai.com/models/1486143/flluxdfp16-10steps00001?modelVersionId=1681047

And this is the one I use with my humble GTX 1070:
https://huggingface.co/ElGeeko/flluxdfp16-10steps-UNET/tree/main

Thanks to the person who made this version and posted it in the comments!

This model halved my render time — from 8 minutes at 832×1216 to 3:40, and from 5 minutes at 640×960 to 2:20.

This post is mostly a thank-you to the person who made this model, since with my card, Flux was taking way too long.

https://redd.it/1l6l4t4
@rStableDiffusion

Good formula for training steps while training a style LORA?

I've been using a fairly common Google Collab for doing LORA training and it recommends, "...images multiplied by their repeats is around 100, or 1 repeat with more than 100 images."

Does anyone have a strong objection to that formula or can recommend a better formula for style?

In the past, I was just doing token training, so I only had up to 10 images per set so the formula made sense and didn't seem to cause any issues.

If it matters, I normally train in 10 epochs at a time just for time and resource constraints.

Learning rate: 3e-4

Text encoder: 6e-5

I just use the defaults provided by the model.

https://redd.it/1l6m1oa
@rStableDiffusion

inference.sh getting closer to alpha launch. gemma, granite, qwen2, qwen3, deepseek, flux, hidream, cogview, diffrythm, audio-x, magi, ltx-video, wan all in one flow!
https://redd.it/1l6q4mm
@rStableDiffusion

I accidentally discovered 3 gigabytes of images in the "input" folder of comfyui. I had no idea this folder existed. I discovered it because there was an image with such a long name that it prevented my comfyui from updating.

many input images were saved. some related to ipadapter. others were inpainting masks

I don't know if there is a way to prevent this

https://redd.it/1l6p6rb
@rStableDiffusion

Any step-by-step tutorial for video in SD.Next? cannot get it to work..

I managed to create videos in SwarmUI, but not with SD.Next. Something is missing and I have no idea what it is. I am using RTX3060 12GB on linux docker. Thanks.

https://redd.it/1l6tfjx
@rStableDiffusion

Best way to animate emojis?

I tried Framepack, but the results were pretty meh. Does anyone know a good method to animate emojis?

https://redd.it/1l70x3s
@rStableDiffusion

Best way to animate an image to a short video using AMD gpu ?
https://redd.it/1l6zoxo
@rStableDiffusion

How to prevent style bleed on LoRA?

I want to train a simple LoRA for Illustrious XL to generate characters with four arms because I've tried some similar LoRAs and at high weight they all have style bleed on the generated images.

Is this a Dataset issue? Should I use different style images when training or what?

https://redd.it/1l72oei
@rStableDiffusion

BAGEL in ComfyUI | All-in-One AI for Image Generation, Editing & Reasoning
https://youtu.be/QRwecGQTev0

https://redd.it/1l7535b
@rStableDiffusion

What is the best solution for generating images that feature multiple characters interacting with significant overlaps, while preserving the distinct details of each character?

Does this still require extensive manual masking and inpainting, or is there now a more straightforward solution?

Personally, I use SDXL with Krita and ComfyUI, which significantly speeds up the process, but it still demands considerable human effort and time. I experimented with some custom nodes, such as the regional prompter, but they ultimately require extensive manual editing to create scenes with lots of overlapping and separate LoRAs. In my opinion, Krita's AI painting plugin is the most user-friendly solution for crafting sophisticated scenes, provided you have a tablet and can manage numerous layers.

OK, it seems I have answered my own question, but I am asking this because I have noticed some Patreon accounts generating hundreds of images per day featuring multiple characters doing complex interactions, which appears impossible to achieve through human editing alone. I am curious if there are any advanced tools(commercial models or not) or methods that I may have overlooked.

https://redd.it/1l75afz
@rStableDiffusion

A Time Traveler's VLOG | Google VEO 3 + Downloadable Assets

https://redd.it/1l788rq
@rStableDiffusion

About 5060ti and stabble difussion

Am i safe buying it to generate stuff using forge ui and flux? I remember when they came out reading something about ppl not being able to use that card because of some cuda stuff, i am kinda new into this and since i cant find stuff like benchmarks on youtube is making me doubt about buying it. Thx if anyone is willing to help and srry about the broken english.

https://redd.it/1l7a9k3
@rStableDiffusion

How to make similar visual?

https://redd.it/1l7bw5n
@rStableDiffusion

HeyGem Lipsync Avatar Demos & Guide!
https://youtu.be/Lefc84zlroA

https://redd.it/1l75bso
@rStableDiffusion

Framepack Studio: Exclusive First Look at the New Update (6/10/25) + Behind-the-Scenes with the Dev
https://youtu.be/hUvZ9VR-9_8

https://redd.it/1l7eug0
@rStableDiffusion

5070 ti vs 4070 ti super. Only $80 difference. But I am seeing a lot of backlash for the 5070 ti, should I getvthe 4070 ti super for $cheaper

Saw some posts regarding performance and PCIe compatibility issues with 5070 ti. Anyone here facing issues with image generations? Should I go with 4070 ti s. There is only around 8% performance difference between the two in benchmarks. Any other reasons I should go with 5070 ti.

https://redd.it/1l7eva5
@rStableDiffusion

Explaining AI Image Generation

Howdy everybody,

I am college professor. In some of my classes we're using ai image generation as part of the assignments. I'm looking for a good way to explain how it works and I want to check my own understanding ai image generation. Below is what I have written for students (college level). Does this all check out?

So how exactly does this work? What is a prompt, what does it mean for an AI to have been trained on your work, and how does an AI create an image? When we create images with AI we’re prompting a Large Language Model (LLM) to make something. The model is built on information called training data. The way the LLM understands the training data is tied to concepts called the Deep Learning system and the Latent Space it produces. The LLM then uses Diffusion to create an image from randomized image noise. Outside of image making we interact with AI systems all of the time of many differing kinds. We usually are not aware of it.

When you prompt an AI you are asking a Large Language Model (LLM) to create an image for you. A LLM is an AI that has been trained on vast amounts of text and image data. That data allows it to understand language and image making. So if something is missing from the data set or is poorly represented in the data the LLM will produce nonsense. Similarly crafting a well made prompt will make the results more predictable.

The LLM’s ability to understand what you are asking is based in part on the way you interact with it. LLMs are tied to an Application Program Interface (API). For example the chat window in Midjourney or Opensea’s ChatGPT. You can also have more complex APIs like Adobe’s Firefly or Diffusionbee (a Stable Diffusion API) that in addition to text prompting include options for selecting styles, model, art Vs photography, etc.

Training data sets can be quite small or quite large. For most of the big name AI models the training data is vast. However you can train AI on additional smaller data sets called Low-Rank Adaption(LoRa) to be especially good at producing images of a certain kind. For example Cindy Sherman has been experimenting with AI generation and may have trained a LoRa on her oeuvre to produce new Cindy Sherman like images.

The training data can be Internet text forums, image forums, books, news, videos, movies, really any bit of culture or human interaction that has been fed into it. This can be much more than what is available on the open Internet. If something exists digitally you should assume someone somewhere has fed it or will feed it into a training data set for an LLM. This includes any conversations you have with an AI.

When something is used to train an LLM it influences the possible outcome of a prompt. So if as an artist your work features praying mantises and someone prompts for an image of a mantis your work will influence the result produced. The AI is not copying the work. The randomness in the diffusion step prevents copying though through concise prompting a very strong influence can be reflected in the final image.

In order for the AI to make sense of the training data it is ran through a Deep Learning system. This system identifies, categorizes, and systematizes the data into a Latent Space. To understand what this means let’s talk about what a digital image actually is. In the digital environment each image is made up of pixels. A pixel is a tiny square of light in a digital display that when combined with other squares of light make up an image. For example the images in this show started as 1792x2668 pixels in size (I later upscaled them for printing). Each of these squares can be one of 16,777,216 color values.

In the deep learning system the AI learns what pixel values and placement that are usually associated with something, for example a smiley face. This allows the LLM to create a latent space where it understands what you mean by a smiley face. It would know what a smile is by data tied to smiling emojis, pictures of people or animals smiling, children’s

drawings, and so on. It would associate faces with human and animal faces but also the face of a cliff or maybe Facebook. However a ‘smiley face’ usually means an emoji so If I asked for a smiley face the LLM would probably give an emoji.

Finally we get to Diffusion. You can think of the latent space as labeled image noise (random pixels) in a great big soup of image noise. In the latent space the LLM can draw out from that noise images based on what it knows something should look like. As it draws the image further out of the noise more detail emerges.

Let’s simplify this process with a metaphor. Let’s say you have a box full of dice where half of the sides are painted black and half are painted white (2 possible colors instead of 16+million). The box holds enough dice that they can lay flat across the bottom of the 400 dice by 600 dice. You ask a scientist to make a smiley face with dice in the box. The scientist picks up the box and gives it a good shaking randomizing the placement of dice. For the sake of the metaphor imagine that all of the dice fall flat and fill out the bottom of the box. The scientist looks at the randomly placed dice and decides that some of them are starting form a smiley face. They then glue those dice to the bottom of the box and give it another shake. Some of the dice compliment the dice that were glued down in forming a smiley face. The scientist then glues those dice down as well. Maybe some of the originally glued down ones do not make sense anymore, they are broken off from the bottom of the box. They repeat shaking and gluing the dice down until they have a smiley face and all of the dice are glued to the bottom. Once they are all glued they show you the face.

In this metaphor you are prompting the scientist for a smiley face. The scientist knows what a smiley face is from their life experience (training data) and conceptualizes it in their mind (latent space). They then shake the box creating the first round of random shapes in the box (diffusion). Based on their conceptualizing of a smiley face they look for that pattern in the dice and fix those ones in place. They then continue to refine the smiley face by continuing to shake and glue dice in place. When done they show you the box (the results). You could further refine your results by asking for a large face or a small face or one off to the left and so on.

Since the dice are randomized it is extremely unlikely that any result will perfectly match another result or that it would prefectly match a smiley face that the scientist had seen in the past. However since there is a set number of dice there is a set number of possible combinations. This is true for all digital art. For an 8 bit image (the kind made by most AI) the number of possible combinations is so vast the likelihood of producing exactly the same image is quite low.1

https://redd.it/1l7hlyk
@rStableDiffusion