Native speedup ('turbo') lora?
Hi @lodestones , first of all, thank you for such a great model, this should be definitely a new standard for 'flux+danbooru' capability.
Do you have any plans on providing 'native' speedup Loras? I've tried different ones for Flux, but they tend to change style and/or other aspects too much. Like '8 step lora', '12 step lora' and so on?
I've got 4070 12GB, and with added CFG it is already twice as slow compared to Flux... Not that it is unexpected, but I'm open to any suggestions how I can speed it up.
Thanks again!
There are a couple here, I haven't tested them yet though
https://huggingface.co/silveroxides/Chroma-LoRA-Experiments
Thanks, already tried these before, I think I managed to make a combo which produces nice photorealistic results with just 20 steps (using Detail Daemon to add details, with dpmpp_2m / sgm uniform).
It's exciting to compare prev / latest versions of this model to witness it converging better and better with every next version, though it still often produces anatomical errors (which often can be fixed with negative prompt adjustments)
Okay, managed to install triton + sage attention under Windows, now speed is fine
Okay, managed to install triton + sage attention under Windows, now speed is fine
What speeds are you getting? Are you using torch compile? Can you perhaps share a workflow?
Here's an image on Civitai, workflow is inside: https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/83eb4411-4d19-4608-ae73-3b6b7d327b14/original=true,quality=90/rgthree.compare._temp_glcgi_00025_.jpeg
the version 29.5 should have native speedup and can generate coherent stuff under 12 steps
@lodestones , I was blown away when switched from v29 to v29.5 yesterday - so basically 30 steps turned into 20 steps, no more any 'accelerator' Loras required. With torch.compile and 1.2s/it it is TOTALLY manageable now.
You can quickly find a good seed for a particular gen, then play with steps count if needed, I usually try adding some 5 steps sometimes.
the version 29.5 should have native speedup and can generate coherent stuff under 12 steps
Is there any possibility of even a small changelog somewhere? I've been enjoying Chroma immensely, downloading each new version as they appear (although I have to wait for either the silveroxides/Chroma-GGUF or Clybius/Chroma-fp8-scaled versions due to limited VRAM).
When I saw v29.5 with no explanation of what the .5 meant, I hoped to see something somewhere mentioning it. But nothing so far until this comment.
Maybe this is all discussed at length in Discord, but that is not an option for me. (They want my phone number, which is a big no from me. Plus any time I've tried to register an account on the off-chance they won't ask for my phone number, the verification link leads to a non-loading page. Plus, I us a VPN, which I won't disable if they'd be asking.)
All I really want is a little changelog somewhere if there's anything important to know.
Thanks.