A look at the NEW ComfyUI Samplers & Schedulers!

Nerdy Rodent
29 Jun 202413:53

TLDRThis video explores new additions to the ComfyUI toolkit, focusing on samplers and schedulers for improving UI design. It tests various samplers like Oiler Plus+ and IP NDM on a model, comparing their performance at different steps and guidance scales. The video highlights the effectiveness of each sampler in generating high-quality images, discussing the impact of changing parameters and the potential artifacts. It also delves into the use of GITS scheduler, noting its performance at varying resolutions, and concludes with a preference for certain samplers over others in higher resolutions.

Takeaways

  • 💡 New features have been added to ComfyUI, including new samplers and schedulers.
  • 📄 Links to relevant research papers are available for those interested in technical details.
  • 🔍 The CFG Plus+ sampler is designed to reduce artifacts such as extra details in images.
  • 👾 The video uses various samplers to generate images of happy aliens, comparing their outputs.
  • ⚖️ Oiler Plus+ can produce burnt-looking images if the guidance scale isn't adjusted properly.
  • 🧬 The ancestral version of Oiler Plus+ shows more differences with higher step counts.
  • 🔄 IP NDM produces different results when steps and guidance scales are adjusted.
  • 📉 Lower guidance scales can result in completely different yet decent images.
  • ⏱️ Deis is effective with fewer steps, producing high-quality images quickly.
  • 🎨 Higher resolutions challenge some samplers, like Gits, which performs well at lower resolutions but not as well at 2K.

Q & A

  • What new features have been added to ComfyUI for UI and sampling?

    -New features added to ComfyUI include a git schedule and a variety of new Samplers, with links to the papers provided for further reading.

  • What is the purpose of the CFG Plus+ sampler?

    -The CFG Plus+ sampler is designed to help reduce problematic artifacts such as extra tals in the image generation process.

  • How does the guidance scale affect the output of the samplers?

    -The guidance scale can significantly affect the output quality. Setting it too high can result in a 'burnt' looking output, while adjusting it can lead to different visual outcomes.

  • What does the script suggest about the ancestral version of Oiler Plus+?

    -The ancestral version of Oiler Plus+ produces images with similar quality to the baseline but may show more differences when the step count is changed.

  • How does the IP NDM sampler behave with different step counts?

    -The IP NDM sampler shows significant changes in the image when the step count is increased, such as from 20 to 40 steps.

  • What is the role of the diffusion exponential integrator sampler (DEIS)?

    -The DEIS sampler is efficient, requiring only 20 steps to produce an image, and can maintain quality even with fewer steps.

  • What is the significance of the git scheduler in the script?

    -The git scheduler is a tool that can perform tasks in fewer steps compared to other samplers, and it allows for adjustments in step count and guidance scale.

  • How do the different versions of Oiler CFG Plus+ sampler compare in terms of artifact reduction?

    -The regular and alternative versions of Oiler CFG Plus+ sampler aim to reduce artifacts, with the regular version showing more human-like features and the alternative version maintaining the quality with different settings.

  • What challenges does the script highlight when using high resolutions like 2K with certain samplers?

    -At higher resolutions like 2K, some samplers, particularly gits, may not perform as well, leading to artifacts and a breakdown in image quality.

  • What is the overall conclusion of the script regarding the performance of different samplers?

    -The script concludes that while some samplers perform well at lower resolutions, they may not maintain the same quality at higher resolutions like 2K, and different samplers have their own optimal settings for best results.

Outlines

00:00

🛠️ New UI Features and Samplers Overview

The script introduces a variety of new features and samplers added to the UI. It discusses the git schedule, new samplers, and the availability of research papers. The author tests these features using the pixart Sigma 1K model, creating happy alien images to demonstrate the capabilities of different samplers. The baseline sampler is compared with the new Oiler Plus+ sampler, highlighting the impact of guidance scale adjustments on image quality. The ancestral version of Oiler Plus+ and its variations with step count changes are also examined, showing distinct differences in image outcomes.

05:02

🔍 Comparative Analysis of Samplers and Gits Scheduler

This paragraph delves into a comparative analysis of different samplers, including IP NDM, ipnd dmore V, and the diffusion exponential integrator sampler (deis), in conjunction with the gits scheduler. The author explores the effects of varying steps and guidance scales on image quality, noting the reduction of artifacts with certain sampler configurations. The use of gits with 10 and 50 steps is evaluated, with a focus on the trade-off between image quality and generation time. The paragraph also touches on the two versions of Oiler CFG Plus+ samplers, their performance with and without gits, and the impact of dynamic thresholding on the final image.

10:03

🖼️ Evaluating Samplers at High Resolutions and Conclusion

The final paragraph assesses the performance of various samplers at higher resolutions, specifically with the Pix art 2K model. The author examines how the samplers handle complex images, such as Cthulhu standing over a kitten in an oil painting style, at increased resolutions. The results indicate that while some samplers maintain quality, others, like gits, struggle with higher resolutions, leading to artifacts. The author also experiments with classifier-free guidance to improve color but notes persistent issues with image artifacts. The video concludes with a summary of findings and a reminder that Patreon members have access to pre-made workflows, encouraging viewers to explore the new features themselves.

Mindmap

Keywords

💡ComfyUI

ComfyUI refers to a user interface design philosophy that prioritizes ease of use and comfort for the user. In the context of this video, ComfyUI is likely a brand or system being discussed, which has introduced new samplers and schedulers to improve user experience. The script mentions 'ComfyUI Samplers & Schedulers' as the main topic of the video.

💡Samplers

In the field of digital imaging and AI, samplers are algorithms that determine how to generate an image based on given parameters. The script discusses various new samplers such as 'Oiler Plus+' and 'Ancestral Oiler Plus+', which are used to create images with different qualities and characteristics.

💡Schedulers

Schedulers, in the context of AI and image generation, are mechanisms that control the sequence and timing of operations in the image generation process. The video mentions 'gits' as a scheduler, which is used to manage the generation process and achieve different results.

💡Guidance Scale

Guidance scale is a parameter in image generation that influences the degree to which the AI adheres to the input parameters. The script shows how varying the guidance scale can lead to different outcomes, such as a 'burnt looking output' or a 'more humanlike' appearance in the generated images.

💡Artifacts

In digital imaging, artifacts refer to unwanted visual elements or distortions that appear in an image. The script discusses how certain samplers like 'Oiler CFG Plus+' are designed to reduce problematic artifacts such as 'extra tals' in the generated images.

💡Steps

Steps in the context of image generation refer to the number of iterations or stages in the process. The script compares different numbers of steps, such as 20, 50, and 10, to demonstrate how the number of steps can affect the quality and appearance of the final image.

💡IP NDM

IP NDM appears to be a specific type of sampler or algorithm mentioned in the script. It is used to generate images, and the script shows how varying the steps and guidance scale can change the appearance of the generated images, such as making the eyes 'absolutely massive'.

💡Diffusion Exponential Integrator Sampler

This term seems to refer to a specific type of sampler that uses a diffusion and exponential integration process to generate images. The script suggests that this sampler is efficient, requiring only 20 steps to achieve a satisfactory result.

💡Gits

Gits is mentioned as both a scheduler and a sampler in the script. It is used to control the image generation process, and the video compares the results of using gits at different steps and guidance scales, noting that it may not perform as well at higher resolutions.

💡Resolution

Resolution in digital imaging is the measure of the detail an image can capture, usually expressed in pixels. The script discusses testing the samplers at different resolutions, such as 1K and 2K, to see how they perform when generating high-detail images.

💡Classifier Free Guidance

Classifier Free Guidance is a technique mentioned in the script that seems to be used to improve the color and quality of the generated images. However, the script notes that even with this technique, some artifacts still remain in the images at higher resolutions.

Highlights

Introduction of new ComfyUI Samplers and Schedulers for UI improvements.

Availability of links to academic papers for detailed understanding.

Inclusion of CFG Plus+ sampler to reduce artifacts like extra tals in images.

Demonstration of using Pixart Sigma 1K model to test new Samplers.

Comparison of different Samplers using the same guidance scale.

Impact of guidance scale variation on image quality and output.

Exploration of Ancestral version of Oiler Plus+ and its effect on image differences.

Adjustment of step count and its influence on image quality.

IP NDM sampler's behavior at different steps and guidance scales.

The peculiarities of using a low guidance scale with IP NDM.

Introduction of ipnd dmore V sampler and its visual outcomes.

Effect of guidance scale on image quality with deis diffusion exponential integrator sampler.

GITS scheduler's performance in generating images with fewer steps.

Comparison of custom sampler Oiler CFG Plus+ with standard versions.

Analysis of the impact of GITS scheduler on image generation.

Evaluation of different versions of Oiler CFG Plus+ with GITS.

Performance of IP NDM with GITS at varying steps and guidance scales.

Assessment of ipnd dmore V with GITS and its image clarity.

Comparison of deis s with GITS at different steps for quality assessment.

Experimentation with dynamic thresholding and guidance scale adjustments.

Use of 2K resolution to test the limits of Samplers and their artifact management.

Summary of Pix art 2K results and Sampler performance at high resolution.

Final thoughts on the effectiveness of different Samplers and GITS at various resolutions.