Furkan Gözükara
patreon
Sampling method DPM++ 2M SDE vs DPM++ 3M SDE with Schedule type Uniform vs Karras vs Exponential
🕑 Added 2024-05-18 19:58:27 +0000 UTCSampling method DPM++ 2M SDE vs DPM++ 3M SDE with Schedule type Uniform vs Karras vs Exponential — I think winner is DPM++ 3M SDE Uniform — Tested on fairly overtrained DreamBooth RealVisXL V4.0 based SDXL model
I wanted to the test newest Sampling method and Schedule types
I think my new favorite is DPM++ 3M SDE Uniform
The tests are concluded on a fairly overtrained RealVisXL V4.0 based Kohya DreamBooth model
Kohya recently made some changes so my older LR values are now causing overtraining. But I wanted to see how well the model can follow my prompts even though being overtrained
Please let me know which one you think best
Not cherry picked results
40 steps and 7 CFG used
After Detailer (ADetailer) is used with 70 steps, prompt: “photo of ohwx man” , 0.5 denoise
One example full metadata (only this one has negative prompt others don’t)
muscular photo of ohwx man wearing an amazing elven steel with gold robe over it full body full pant armor in a garden in ancient rome empire, hd, hdr, uhd, 2k, 4k
Negative prompt: nipples , nipple
Steps: 40, Sampler: DPM++ 2M SDE, Schedule type: Uniform, CFG scale: 7, Seed: 2129378110, Size: 1024x1024, Model hash: d4ca40b15c, Model: 15_may_L2_snr_8e_6, ADetailer model: face_yolov8n.pt, ADetailer prompt: photo of ohwx man, ADetailer confidence: 0.3, ADetailer mask only top k largest: 1, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.5, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer use separate steps: True, ADetailer steps: 70, ADetailer version: 24.4.2, Version: v1.9.3
Comments
Furkan Gözükara
didn't find best yet but reducing LR to 0.000008 (8e-6) (previously 0.00001 - 1e-5) helps greatly so far. I am still researching
Pew
"Kohya recently made some changes so my older LR values are now causing overtraining." Doc, having you in our corner is the next best thing to my being not entirely dumb. :) Do you know yet, the fix, or corrected LR value to enter into your existing (most recent) Kohya 24GB VRAM script?
