This series of LoRA models are extractions from chroma-unlocked-v47-flash-heun and despite the indicating name it is not restricted to be used with Heun sampler nor that lowest step is supported across all ranks.
I can't upload every model in one go so for any of the other ranks in the meantime, visit this huggingface model repo
Description
Rank 64 can be used with very low CFG below 1.3 and will allow for significant reduction of step count depending on sampler/scheduler used. It can also be used without CFG (set to 1.0 in ComfyUI) at slightly higher step than with CFG.
FAQ
Comments (13)
What does this lora do exactly ?
I am guessing faster generation...
https://www.reddit.com/r/StableDiffusion/comments/1o3revh/chromaflashhuen_loras_are_now_available_on/
@CyclopsGER ok im seeing results. it actually make my images look way better and realistic. im using the rank 256. without 8 steps
I've never heard about the rank influencing the ideal CFG. Is that the case for other models "lightning" loras as well?
Can't answer that really as I don't know any model that has a lightning variant that has similar architecture as Chroma
@S1LV3RC01N ,, thanks for these,, would you recommend testing out rank 64 and 256, or different ones? I'm new to using lightning (flash) loras with Chroma. Thanks
Does negative prompt still work with CFG set to 1.0?
No. CFG set to one is like disabling CFG.
If you really need to use negative prompting and still want the speed benefit of CFG 1.0, you can try comfyui-adaptive-guidance, and get something in between.
Not sure why folks are saying Negative Prompts don't work with CFG 1.0, as all of my images use CFG 1.0 with Negative Prompts working well to clean up details. I'm using EulerCFG++ & EulerAncestralCFG++ with the CFGGuilder node in ComfyUI.
@NeonScreams Because programmatically, CFG is disabled when set to 1.0 in ComfyUI. When CFG is disabled, you do not get negative conds fed to model unless using something like NAG. So your experience is most likely either imagined or you are using something else that negates CFG disabling and as such you are running at same speed as if you used CFG anyway.
@S1LV3RC01N Before I discuss, I recognize you and your contributions to Chroma, both here and HuggingFace; So I believe you when you say that is how Negative Prompts + 1.0CFG is being handled in ComfyUI. I can only wonder if the CFGGuider(node) + SamplerEulerCFG++(node) + OptimalStepsScheduler(node) + SamplerCustomAdvanced(node), have internal NAG without exposing it to the front end. I don't program, and while I can get the gist of the Python in nodes I haven't had a reason to go look.
But I can say with confidence that as I use Chroma1-HD as my primary model with all sorts of LoRAs, Positive Prompts and Negative Prompts, I'm seeing a noticeable 'avoidance' of undesired traits in the images produced. "Anime, Cartoon, Toon, Comic", for instance have an immediate noticeable effect on the outcome.
In this case I must be seeing the aforementioned Adaptive Guidance with the ComfyUI nodes I'm using, without having realized it.
there is a new type of guidance exactly for this case:NAG(normalized attention guidance)
searchable in comfyui manager in a pack of other alternative guidance methods called sd-perturbed-attention.