SD3.5 Med/Lite (11GB) Improved Dual CLIP
Guide to picking SD3 model (If you can run FP16 then run FP16 over FP8)
If you have a 3090/4090 I suggest the FP16 Hybrid model
If you have a older card I suggest the LARGE FP8 model
If you have a 8GB card I suggest this model unless you don't mind a wait.
Works in Comfy-UI without any modification just load checkpoint and go.
Medium Model SD 3.5 Base
UNET is not modified other then to quantize to FP8
BF16 T5xxl (From FP32), and Improved CLIP-L
CLIP-G has been removed for this version in Dual CLIP version. (Reduced size and increased IT's with little loss in quality in testing)
My IT's per Second on an old 3050 8GB RTX
SD 3.5 Medium (Dual CLIP Hybrid)
8GB - Full FP8 Model lost to much quality. It was 6.9GB but I considered it unusable (Wait for GGUF or NF4)
11GB = 1IT per second
SD 3.5 Large (Triple CLIP FP8)
13.5GB = 6-8 Seconds Per IT
22GB Hybrid = 6-8 Seconds Per IT
26GB (FP16 FULL) = 10-15 Seconds per IT