Maybe you feel the same way as me...I am always interested in new things and especially in new models here on CIVITAI.
Therefore, I would like to share my first experiences in dealing with SD3.5 and hope to stimulate a discussion.
I work with Stable Diffusion only as a hobby and therefore the following text should not be understood as a scientific evaluation.
Hardware:
System: 64-Bit, Win 11
Prozessor: AMD Ryzen 7 5800X 8-Core, 3.80 GHz
RAM: 64 GB
GPU: NVIDIA GeForce RTX 4060 Ti 16 GB
Tool: WebUI Forge:
To find good results for SD3.5 I experimented with different settings of CFG scale and sampling steps. The included images were created without LoRa.
Checkpoint: STOIQO NewReality SD3.5 PreAlpha (https://civitai.com/models/161068/stoiqo-newreality-flux-sd35-sdxl-sd15)
VAE / Text Encoder: clip_g + clip_l + t5xxl_fp16
Sampling method: Euler
Schedule type: Simple
Sampling steps: 20
Distilled CFG Scale: 0
Seed: 1891278805
Resolution: 896 x 1152
CFG Scale:
From CFG Scale 8 onwards, clear changes can be seen. CFG Scale 10 and above are no longer usable. I find the results in the range of CFG 3 - 7 to be the best.
computing time:
In the images above you can see how the quality and calculation time of the results changes.
I think it is a good solution to create test calculations with 10 steps. In the images with 10 steps you can see everything you want to see. If you have found one or two good results, then improving the quality by recalculating with around 30 - 40 steps is recommended.
The prompt for the image above is a text-based description.
Similar to a flux model, I also achieved good results when the description consists of individual tokens. (see the next image)
I would be very happy if you share your experiences in the comments. If you have any suggestions on how I can improve my work with SD3.5, know of better settings or other things that can improve the results even further, please let me know.
My personal conclusion:
In my opinion, SD3.5 is of very good quality. It also doesn't produce as many errors (malformations) as with Flux. When working with Flux, I usually only get 25% good results. (But maybe that's due to the input prompts or the models used)
Due to the long computing times (similar to Flux), I probably won't choose SD3.5 as the best option for me any time soon. I will definitely continue to work with SDXL and occasionally with SD1.5.
Installation instructions for SD3.5 for Forge WebUI can be found at: https://www.reddit.com/r/StableDiffusion/comments/1ge7kpu/sd35_model_on_webui_forge/