Do you teste it? how to get good quality?
Do you have a workflow? due to TencentARC haven't share the infernece code and workflow now, I had use many method to try the model like Flux.1, but can't get good quality. if you have sample workflow, can share to me? thanks!
At the same time, I want to share my model to you, I think it was a very good base Flux model, only 6-10 step, can get very good quality and detail.
https://huggingface.co/wikeeyang/Flux.1-Dedistilled-Mix-Tuned-fp8
I have tested it, but it's subpar with Flux Dev or schnell. Text was gutted from it as it can't do text, doesn't do human body well. The image should have my workflow, but my workflow is massive, not beginner friendly either. I don't have a lot of high hopes for it, but good for those with low end gpu
Much thanks for share the test result and workflow to me! my result as like you had got, waiting for the official script and workflow, also no more high hope for it.
if use Xlab's Sampler, the speed and the LoRA compatibility is not so good. Much thank a lot again...
I read on civitai to use the SharkSampler instead of Ksampler, but I have no idea where the SharkSampler comes from
Yes! I had tested workflow from Civitai yet yesterday, can get good result, but too slow...in this case, there is still no advantage for reserch...
Tencent ARC Lab closed the model page. I tested their pruned model in chunks, and I think they used a wrong strategy. Since the main data weights for FLUX single and double blocks are mostly distributed at the ends, ARC's simple truncation of the tail caused severe loss of double block data, leading to very high error rates.
Oh, interesting on the closed model part and I know text was impossible for it from my tests. Guess we wait and see what they come up with next.