BigAsp3 is going to be a big hit.
I’ve also done a few tests of my own and I can assure you that this model is very robust, which is all the more surprising and promising for a version that isn’t yet finished. It adheres very well to the prompts and there are FAR fewer anatomical errors. I can’t wait to see what the final version will be like.
got any advice for sampler settings and such in this early stage?
It should behave similar to other Flux 2 Klein 9B checkpoints, so generally just Euler is enough. I've been running at Euler 40 steps, cfg=5. The ODE samplers at 20 steps are an okay alternative but I didn't see any benefit in my initial testing. UniPC is also an option.
At this point the biggest gains in a workflow will come from the prompt, since bigASP 3 was trained exclusively on long, detailed prompts. It has trouble with short, tag spammy prompts compared with long, detailed prompts. Your LLM of choice should do a decent job "upgrading" prompts until a dedicated prompt enhancer is trained.
got any advice for sampler settings and such in this early stage?
Having carried out some in-depth testing on the latest release of BigAsp3, I’ve achieved good results from 25 to 30 steps.
I’m also using this distilled LORA model (r256) at arround 0.4 in strenght : https://huggingface.co/vafipas663/flux2-klein-base-9b-distill-lora/tree/main
I use DDIM, but Euler is indeed sufficient, in conjunction with the Flux2Scheduler.
CFG 8 (I’ve noticed better adherence to prompts with a higher CFG).
As @fancyfeast points out, BigAsp3 really likes fairly long and precise prompts, so using an LLM (personally, gemma4 31b in nvfp4 https://huggingface.co/llmfan46/gemma-4-31B-it-uncensored-heretic ) with a system prompt (natural language-oriented for klein9b) that generates ~200-words prompts will yield high-quality outputs.
As for the maximum output resolution, it’s the same as with the stock klein9b: beyond 2000px, the quality deteriorates, so I often stick to something like 900x1500.