title: Multi-Resolution Sampling/Scheduling Improves Image Consistency
author: yoinked
tags: image, diffusion, sampling, scheduler, sampler
abstract: >-
sampling at multiple resolutions in the sampler improves outputs at
out-of-distribution resolution, effectively doubling
date_published: 2024-09-24T00:00:00.000Z
paperid: 2024-0924.1
Multi-Resolution Sampling/Scheduling Improves Image Consistency
1. Abstract
In this paper, we will present a novel technique for sampling diffusion models, in which the model's effective resolution can be doubled or even tripled. We also present a new series of schedulers which, can improve sampling efficiency by enhancing the sampler's strength in the middle steps of the generation process. These tools are presented in code samples available here.
2. Methodology
2.1 Sampler
The sampler we present is based on DPM++ 2S Ancestral, with 2 modifications; the first is that we implement CFG++, which enables us to use a lower Classifier Free Guidance value for inference, which leads to better prompt adherence; the second modification is our novel method of improving the model's consistency at higher resolutions. The method involves down-sampling the base latent [denoted as Lin
], which, while it is a "lossy" procedure, still enables for the model to understand those usable latents. Afterwards, we pass the down-sampled latent [denoted LDW
] through the diffusion model, which, in higher resolution cases, will result in a more consistent latent, due to the model having been trained on that lower resolution. After the first model call, we up-sample LDW
and merge it with Lin
at a ratio of M
, which is simply defined as the root of the current step number divided by the total number of steps. Once the latents are merged [denoted as LM
], we pass that into the diffusion model again, to get rid of the up-sampling abnormalities and to refine it at the higher resolution whilst having a consistent base. [All denoted in the diagram below.]
2.2 Scheduler
The schedulers we present help to assist the sampler mentioned above, by providing with a higher sigma value during middle steps, effectively increasing the strength of the model, whilst retaining the consistency of the multi-res sampling. The following 3 schedulers are the best schedulers for this sampler we've tested:
Sinusoidal scheduler with scaling factor (sinsf)
This scheduler is the "angriest" scheduler out of the 3, in other words, the scheduler that decreases the fastest at the start. While this attribute may sound un-optimal, some popular schedulers are also angry, most notably, the exponential scheduler. The scheduler is defined as ( ( ( sigma_min + ( sigma_max - sigma_min ) * ( 1 - sin( π / 2 * x ) ) ) / sigma_max ) ** sf ) * sigma_max
, in which x
is defined as a list of numbers from [0 - 1] (inclusive) of length N
, which is the number of steps, and in which sf
is equal to the scaling factor provided, usually around 3 for sinsf
.
Inverse cosinusoidal scheduler with scaling factor (invsf)
This scheduler is designed as being a mid-step between sinsf
and dynsf
, which attempts to be more like the most used Karras scheduler. The scheduler is defined as ( ( ( sigma_min + ( sigma_max - sigma_min ) * ( 0.5 * ( cos( x * π ) + 1 ) ) ) / sigma_max ) ** sf ) * sigma_max
, in which x
is defined as a list of numbers from [0 - 1] (inclusive) of length N
, which is the number of steps, and in which sf
is equal to the scaling factor provided, usually around 3 for invsf
.
Reactive cosinusoidal scheduler with scaling factor (dynsf)
This scheduler is the "calmest" scheduler out of the 3, in other words, the scheduler that decreases the slowest at the start. The scheduler is defined as ( ( ( sigma_min + ( sigma_max - sigma_min ) * ( cos( x * ( π/2 ) ) ) )/sigma_max ) ** sf ) * sigma_max
, in which x
is defined as a list of numbers from [0 - 1] (inclusive) of length N
, which is the number of steps, and in which sf
is equal to the scaling factor provided, usually around 4 for dynsf
.
3. Results
We tested the sampler by generating images outside of its "attention range" (for sd1.x it is 768x768 pixels), in this example, we ran a generation on an image of size 1024x1792; the sampler we used to compare against was DPM++ 2S Ancestral CFG++, so the only difference in the sampler should be the new internal up-sampling. We used invsf
for our scheduler of choice, as it still helps non-internal samplers.
The deformities (multiple bodies/torsos) in the first image are caused due to the model not capturing enough in its "attention range" for it to understand how the main feature of the image should appear.
4. Conclusion
In conclusion, our proposed new sampler (labled dpmpp_2s_ancestral_cfgpp_intern
) attempts to solve some shortcomings of regular Stable Diffusion models, in which the model "forgets" where parts of the main subject should go. We solved it by down-sampling latents before calling it to the model, this does leave the downside of suffering on smaller resolutions, as we would effectively down-sample too far and be left with near-unworkable latents once we up-sampled. We also presented 3 new schedulers, each with their use-cases, improving image quality on all samplers, but mainly on our proposed sampler.