#Is it possible to get seedream level capability with swarmui?
1 messages · Page 1 of 1 (latest)
We have Flux 2 now and it is miles ahead of Flux kontext. It runs in SwarmUI perfectly fine
@wind nymph oh lol. I thought flux2 was like a flux dev. But it’s more like kontext?
It is a Flux dev replacement, but it also has the capability to be given up to 10 images as references, it does everythyng in one model
@wind nymph Ohhhhh baby that’s Saweet. I mean swarmui made it easy to switch, but all in one gonna be sweet. Thanks for info! 10 images is wild wow
(wild slow too, each input image basically double the gen time)
Yeah Flux 2 is a huge model. So bring lots of VRAM and FLOPS or you will be waiting up to double digit minutes for 1 image in some cases
In bf16 the Flux2 model and its text encoder is around 100 GB to download
Even the fp8 quant does not fully fit into 24GB of VRAM when running, so some amount of swapping is needed. Once you go to GGUF quants it will be even slower due to the extra unpacking overhead
This is why Z-Image stole all of the spotlight from Flux 2. It is really good, yet people can actually run it on their hardware due to being a much smaller model.(but the edit model is not yet released)
Oh man flux2 is out of the question then. That is just too big for my 4070 Super.
So to get back on track, to my original question, to get seedream level capability, where you load in a few images and make changes, do people just load it in init and prompt the changes?
Can you load in multiple photos into the prompt box and prompt changes with that? Or maybe there is instructions somewhere?
you're looking for an Edit model
best one beside Flux2 is Qwen-Edit-2509, it can take up to 4 images input
You can get Flux 2 running on 12GB of VRAM if you really try, but yeah won't be fast.
Qwen Edit is a great edit model that is easier to run(tho still not exactly small), also has lightning LoRAs to make it very fast at 4 step or 8 step generation
Hmmmmm i wasn’t aware of qwen edit. I will try it 😁BTW guys thank you for the support here. Really appreciate it. I got people telling me to use all kinds of things, and I was pretty sure swarm with the right models can do it better especially with comfy backend.
#gens message type of things QE can do with multiples inputs
Interesting I didn’t realize. I should check out this discord more. Lots of great info I think. Thanks for the link!
a 4070 super and 64 gigs of system ram should be able to run flux2 i think
check #announcements
also jsyk Z-Image Edit is coming "soon" and is likely to be faster and better vs qwen, though not sure yet until it's out ofc