I get much better results with Automatic1111 compared to ComfyUI. Can I fix this (and how)?
I consistently get much better results with Automatic1111's webUI compared to ComfyUI even for seemingly identical workflows. The more complex the workflows get (e.g. multiple LoRas, negative prompting, upscaling), the more Comfy results break down.
I've attached some example pictures and also a link to a Miro Board which shows the exact workflows.
Miro: https://miro.com/app/board/uXjVNRna5Hc=/?share_link_id=661735166309
ComfyUI images look much more cartoonish and unrealistic. With LoRas and latent upscaling I get totally corrupted images for the same workflow. Auto1111 holds up just fine.
I have some nice workflows in Auto and would love to replicate them - if possible exactly - in Comfy. From what I understand, exact same parameters should give the exact same image as a result. However, ComfyUI and Automatic1111 seem to be processing Stable Diffusion quite differently.
Does anyone know how I can get my exact Auto1111 results in Comfy?
I noticed that reducing the cfg scale in ComfyUI gets me more pleasing results, but it's still not at all the same as in Automatic1111. And I do wonder why they are so extremely different. (Seeds in Comfy are much longer by default. Does that mean anything?) Any ideas?
Examples: Comfy (left) vs. Automatic1111 (right) (Model: photon, 20 steps, cfg: 6, scheduler: DPM++ SDE Karras/dpmpp_sde karras, seed: 2776553318, upscaling: latent (bicubic) with 10 steps)
You probably applied prompt weighting same as A1111. A1111 tends to have a very weak effect of prompts compared to ComfyUI, so you must have given strong weighting to match it. When convert to ComfyUI, you should lower the weighting.
And don't forget embedding: prefix for embeddings.
You can try my node.
You should stop trying to copy your old A1111 stuff and start from scratch. See if you can replicate the results of other people's workflows first, if not something is wrong with your setup. You can look at example workflows with example outputs here:
https://comfyworkflows.com/
You probably applied prompt weighting same as A1111. A1111 tends to have a very weak effect of prompts compared to ComfyUI, so you must have given strong weighting to match it. When convert to ComfyUI, you should lower the weighting.
And don't forget
embedding:prefix for embeddings.
Thank you so much, that did help a great deal. Quality is now massively improved. I feel I'm not quite at the level of my A1111 workflow yet, but much much closer. I'll continue tweaking it.
You can try my node.
Greatly appreciated!! Your node seems to be exactly what I was looking for. Right now, I'm getting a tuple-error with the note, however. I'll post it there.
You should stop trying to copy your old A1111 stuff and start from scratch. See if you can replicate the results of other people's workflows first, if not something is wrong with your setup. You can look at example workflows with example outputs here:
https://comfyworkflows.com/
Thanks for the resource! My goal is still - in so far it's possible - to get the exact same result first (because I'm already extremely pleased with them) and then work from there to possibly tweak it even more. But you're right, there's much more to learn with Comfy.
Thanks for the resource! My goal is still - in so far it's possible - to get the exact same result first
My suggestion was primarily meant to rule-out a problem with your installation... I think you should consider it. If you can't replicate images that other people have made using their workflows then something is wrong with your installation and you need to start there.
I know this might sound too simple of a fix, but same as you, I literally just switched over and been getting all these contrasty images even though everything setting was the same.
The CFG (ususally on auto1111 I use around 7), I lowered it to a 3 and everything turned out to be the same as Auto1111.
Hope it helps
I am having the same issue and it is rather frustrating. I installed @shiimizu 's node and it got a lot closer, but still the shading and overall composition still just seem a tad worse than Automatic1111 to the point that it is honestly a little detering. I really like the automation of Comfy, but the image quality being worse or feeling a little more off than A11 is holding me back from jumping over completely