RayVietii_DryRender-DD2.7
RayVietii_Distilled-Diffusion
###
DDPM scheduler setting comparison:
RayVietii-DryRender:
βs = 0.0 | βe = 0.0095
Kohya_ss default setting:
βs = 0.00085 | βe = 0.012
Huggingface:
βs = 0.0001 | βe = 0.02
###
built-in positive prompt:
```
good_pose, good_face, good_eyes, good_eyebrows, good_eyelashes, good_nose, good_lips, good_mouth, good_teeth, good_chin, good_cheeks, good_forehead, good_ears, good_hair, good_bangs, good_neck, good_clavicle, good_shoulders, good_shoulder_blades, good_arms, good_elbows, good_forearms, good_wrists, good_hands, good_fingers, good_fingernails, good_thumbs, good_chest, good_breasts, good_waist, good_belly, good_navel, good_back, good_spine, good_ribs, good_hips, good_thighs, good_knees, good_calves, good_ankles, good_feet, good_toes, good_toenails, good_skin, good_proportions, good_anatomy, good_posture, good_body
```
Facts about this model:
Hi! This is Ray Vietii, from what I've gathered from the paper
Judging from the formula, Xt = βs...βt = X0 (simplified), (t1, t2, ..., tn )
> While it may seem complex at first, the process is actually quite straightforward. Here, Xₜ represents the image at timestep t , and Xₜ₋₁ represents the image at the previous timestep. ϵis our randomly generated unit Gaussian noise. Since it is a unit Gaussian, its variance is one. When we multiply it by the term square root of βt, its variance becomes βt. We also scale down Xₜ₋₁ by square root of 1-βt to ensure that the variance of Xₜ does not grow when we add noise. This is essentially a balancing term.
The βt parameter controls the amount of Gaussian noise added to the image. The authors call this the variance schedule, which ramps up at higher values of t. In the original work by [Ho et al. (2020)](https://arxiv.org/abs/2006.11239), betas are put in a linear space from β1=0.0001 to βT=0.02 with T=1000 diffusion steps. They are relatively small compared to the normalized image pixel values between [−1,1].
What does this mean is, it started with no denoising at all, keeping the dataset image as is, which means X1 = X0 (original image). βe = 0.0095, this value is significantly smaller then the recommended 0.02, which means, with 0.0095 β, the dataset image is never truly become a pure noise, which give us an answer of why it's smart at predicting, because this model have no hallucination (pure noise) to start off during reverse process (denoising).
Conclusion: This model having traits as those LCM SD models. It's so sensitive to prompt, especially weighted or emphasized prompt eg (), (a:1.2), it's guessed the image with only CFG 1.0 [Cover Image].
My goal is to push SD1 potentials comparable to those SDXL models, but as we know already, SD1.5 only have 983m parameters, significantly smaller than SDXL which have 3.5b parameters. At least my hope with this model is as forgotten SD2.1.
>That being said, Happy generating, friends! :D
> -Ray Vietii