The average ComfyUI experience when downloading a new workflow
https://redd.it/1oqwj6t
@rStableDiffusion

AMD Nitro-E: Not s/it, not it/s, it's Images per Second - Good fine-tuning candidate?

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

Cathedral (Chroma Radiance)

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

Nvidia cosmos 2.5 models released

Hi!
It seems NVIDIA released some new open models very recently, a 2.5 version of its Cosmos models, which seemingly went under the radar.

https://github.com/nvidia-cosmos/cosmos-predict2.5?tab=readme-ov-file





https://github.com/nvidia-cosmos/cosmos-transfer2.5

Has anyone played with them? They look interesting for certain usecases.



EDIT: Yes, it generates or restyles video, more examples:

https://github.com/nvidia-cosmos/cosmos-predict2.5/blob/main/docs/inference.md

https://github.com/nvidia-cosmos/cosmos-transfer2.5/blob/main/docs/inference.md

https://redd.it/1oqz2xy
@rStableDiffusion

Outdated info on the state of ROCM on this subreddit - ROCm 7 benchmarks compared to older ROCm/Zluda results from a popular old benchmark

So I created a thread complaining about the speed of my 9070 and asked for help with choosing a new Nvidia card. A few people had good intentions but they shared out of date benchmarks using a very old version of ROCm to test AMD GPUs.

The numbers in these benchmarks seemed a bit low, so I decided to replicate the results as best as I could comparing my 9070 to the results from this benchmark:

https://chimolog.co/bto-gpu-stable-diffusion-specs/#832%C3%971216%EF%BC%9AQwenImageQ3%E3%83%99%E3%83%B3%E3%83%81%E3%83%9E%E3%83%BC%E3%82%AF

Here are the numbers I got for Sd1.5 and SDXL, getting them as close as I could to the prompts/settings used in the benchmark above:

SD1.5 512 10 batch 28 steps

- Old 9070 benchmark results 30 seconds
- New rocm 7 9070 13 seconds


On the old benchmark results, this puts it just behind 4070. Further comparison showed the following results for the following GPUs in the old benchmark:

- 8 seconds on 5070ti
- 6.6 seconds on 5080

SDXL 832x2316 28 steps

- Old 9070 benchmark 18.5 seconds
- New rocm 7 9070 7.74 seconds

On the old benchmark results, it's once again just behind 4070. Further comparison showed the following results for the following GPUs in the old benchmark:

- 4.7 seconds on 5070ti
- 3.8 seconds on 5080

Now don't get me wrong, Nvidia is still faster, but, at least for these models, it's not the shit show it used to be.

Also, it's made it clear to me that if I want a far more noticeable performance improvement, I should be aiming for at least the 5080, not the 5070ti, since the difference is about 40% between the 9070 and the 5070ti Vs almost 100% difference between the 9070 and 5080.

Yes, Nvidia is the king and is what people should buy if they're serious about image generation workloads, but AMD isn't as terrible as it once was.

Also, if you have an AMD card and don't mind figuring out Linux, you can get some decent results that are comparable with some of Nvidia older upper mid range cards.

Tldr: AMD have made big strides in improving their drivers/software for image generation. Nvidia still the best though.

https://redd.it/1or6mvu
@rStableDiffusion

Qwen Image Base Model Training vs FLUX SRPO Training 20 images comparison (top ones Qwen bottom ones FLUX) - Same Dataset (28 imgs) - I can't return back to FLUX such as massive difference - Oldest comment has prompts and more info - Qwen destroys the FLUX at complex prompts and emotions

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