Stable Diffusion XL - The impact of CFG Scale

Steps 60, Sampler DPM++ 2M Karras, CFG scale 1-30, Seed 33, Size 1024x1024, Model hash 31e35c80fc, Model sd_xl_base_1.0, Version v1.6.0
1 batch, batch size 4.
Same prompts and negative prompts. Created out of my own curiosity.
#stablediffusion #sdxl #cfgscale #ai

[@SenTenshi]

CFG is tricky and should be adjusted depending on the type of work and model we're using. Especially if we're using less or non convergent samplers such as "ancestral" or "SDE", etc... For photorealism it's usually pretty bad to have high CFG, results would look more 3D than anything else. But art or cartoon, manga, etc... it's different. Add LoRa, Lyco and Neg Embeddings and it will require even more testing and weighting haha ! Loved your dark souls art like prompt !

[@nernberon]

That's right. The model you use also has an impact on how high or low you can scale CFG. Some work better with values like 10+ and need a lot of negative prompts. Newer ones work best with around CFG 6 and need way less neg. prompts. It feels like models like SD XL are smarter (trained better) and only get confused by too many negative prompts. Sampling steps also impact a lot. And the chosen form factor. (1:1, 4:3, 16:9)