lodestones

/

Chroma1-HD

like 306

I'll put together a quick simple workflow for the new settings that should get you up and running successfully.
I'll reply to your original post when finished.

I really don't know what's happened with Chroma since it hit the final epoch. The Comfy workflow doesn't work well for me at all... Silveroxide's and Lodestone's workflows are the same. I've tried countless models — fp8, scaled, gguf and radiance — and the results are always awful. Even at 40-50 steps with CFG ranging from 1 to 8, it's like the image isn't converging. Everything comes out looking half baked, deformed, bad anatomy... The only way to get a remotely acceptable image is to use an accelerator LoRA and that either hugely influences the style, affects prompt adherence or still fails to produce a clear image.

Never used to have these issues with Chroma before the Chroma1-HD release and the new default Comfy workflow. It's really annoying because Chroma was always my go-to and now I just can't seem to get a single model to work properly even with high step counts...

And as for the 'hybrid loader' node from Silveroxide — that seems to lead to even worse performance and longer generation times but with even worse convergence.

Image should contain the workflow with additional notes. Apologies if the content is objectionable, as its how I primarily work with Chroma.

hold pls while I 'Disney' the image

Far more family wholesome. shudders

I really don't know what's happened with Chroma since it hit the final epoch. The Comfy workflow doesn't work well for me at all... Silveroxide's and Lodestone's workflows are the same. I've tried countless models — fp8, scaled, gguf and radiance — and the results are always awful. Even at 40-50 steps with CFG ranging from 1 to 8, it's like the image isn't converging. Everything comes out looking half baked, deformed, bad anatomy... The only way to get a remotely acceptable image is to use an accelerator LoRA and that either hugely influences the style, affects prompt adherence or still fails to produce a clear image.

Never used to have these issues with Chroma before the Chroma1-HD release and the new default Comfy workflow. It's really annoying because Chroma was always my go-to and now I just can't seem to get a single model to work properly even with high step counts...

And as for the 'hybrid loader' node from Silveroxide — that seems to lead to even worse performance and longer generation times but with even worse convergence.

Image should contain the workflow with additional notes. Apologies if the content is objectionable, as its how I primarily work with Chroma.

Thanks @NeonScreams - unfortunately, the image doesn't appear to have come through in your post though.

Some notes on the WorkFlow- Be forgiving, I made it in 3 hours. I'm still working on my primary 'WorkDesk' style workflow for Chroma1-HD.

On this WF you'll see:

• EulerCFG++ at 1.0
  OptimalStepsScheduler 24 steps (15-40 seems the best range

That gets you 1.0 CFG so you can use your LoRA strengths for variations. And a Scheduler that doesn't mind if you add in a Sigma Detailer later in the chain for pulling out better face details.

But in the meantime, this is a solid "It Just Works" configuration.

Personally I agree with Lodestone that using the Custom BETA-Version Scheduler that allows for adjustment of the Alpha/Beta-Sigmas around 0.46 will generally result in a more detailed final image. The difficulty I ran into was adding 3-5 LoRA's, and balancing the CFG & A/B Sigma along with them. Not that it couldn't be done, but it seemed like something I'd save for a 'Final Draft', instead of 'I just want to glance at how these settings turn out'.

I swapped the image for something less XXX. The image is linked at
https://huggingface.co/lodestones/Chroma1-HD/discussions/24?image-viewer=692cd53ce34f1ce85a704ab5-257071A2AD19F9AC6C4CBD769F4328A73291B2F8

I tested to be sure, and it contains the drag-drop workflow. Sorry for the headaches all, hope it helps.

I really don't know what's happened with Chroma since it hit the final epoch. The Comfy workflow doesn't work well for me at all... Silveroxide's and Lodestone's workflows are the same. I've tried countless models — fp8, scaled, gguf and radiance — and the results are always awful. Even at 40-50 steps with CFG ranging from 1 to 8, it's like the image isn't converging. Everything comes out looking half baked, deformed, bad anatomy... The only way to get a remotely acceptable image is to use an accelerator LoRA and that either hugely influences the style, affects prompt adherence or still fails to produce a clear image.

Never used to have these issues with Chroma before the Chroma1-HD release and the new default Comfy workflow. It's really annoying because Chroma was always my go-to and now I just can't seem to get a single model to work properly even with high step counts...

And as for the 'hybrid loader' node from Silveroxide — that seems to lead to even worse performance and longer generation times but with even worse convergence.

Image should contain the workflow with additional notes. Apologies if the content is objectionable, as its how I primarily work with Chroma.

Thanks @NeonScreams - unfortunately, the image doesn't appear to have come through in your post though.

Let me know if that last post has a good link, or if you need some other host for it. (CivitAI or similar)

I have no problems with it, Chroma1 HD works better than v50. I have been using a custom merge, I have no problems at all and it's even better than older versions