Report good images and images to be improved. 【报告优秀和待改进的图像】

[JasonGUTU]

Dear community members,

First of all, I would like to express my sincere gratitude to everyone who uses and supports SUPIR. Your feedback is a key driver for our improvement and development. In order to further improve the performance and user experience of our model, we sincerely invite you to share your experience in this issue.

首先，我想对每一位使用并支持 SUPIR 的用户表示衷心的感谢。您的反馈是我们改进和发展的关键动力。为了进一步提升我们软件的性能和用户体验，我们诚邀您在这个issue中分享您的使用经验。

We welcome your reports:

• Exceptional Images: Share images that you think our software did a great job with. Please briefly describe the image content and why you think the treatment worked well. If possible, include a link to the image or the image itself.
  特别好的图像：分享那些您认为我们软件处理得非常出色的图像。请简要描述图像内容，并说明为什么您认为处理效果好。如果可能，附上图像链接或图像本身。

• Images to be improved: If you encounter some poor image processing, please also report it here. Sometimes, unsatisfactory processing results can be caused by improper use of SUPIR. To help us diagnose the problem more accurately and provide a more effective solution, please describe the problem in as much detail as possible and provide image examples and what you think may be the cause or suggestions for improvement. At the same time, it is very important that in order for our team to better understand the problem and try to reproduce it, we recommend that you also provide the input original image. We understand that every image and every process is unique, and by providing enough information, you will help us deeply understand how SUPIR performs in different situations and optimize accordingly.
  待改进的图像：如果您遇到了一些图像处理效果不佳的情况，请在这里报告。有时候，不理想的处理效果可能是由于软件使用不当导致的。为了帮助我们更准确地诊断问题，并提供更有效的解决方案，请尽可能详细地描述问题，并提供图像示例以及您认为可能的原因或改进建议。同时，非常重要的一点，为了让我们的团队能够更好地理解问题并尝试复现，建议您也提供输入原图。我们理解每一张图像和每一次处理都有其独特性，通过提供足够的信息，您将帮助我们深入了解软件在不同情况下的表现，并作出相应的优化。

We believe that with your valuable feedback, together we can take this project to new heights. Please remember to keep your feedback objective and respectful, we are committed to providing an open and inclusive communication environment.

我们相信，通过您的宝贵反馈，我们可以一起将这个项目推向新的高度。请记得保持反馈的客观和尊重，我们致力于提供一个开放和包容的交流环境。

Thank you for your contribution and support!

best wishes,

XPixel, The Author Group

[FurkanGozukara]

This is a great model. It literally destroys the very expensive SaaS Magnific. I compared with 10 unique images here : #38

My first suggestion would be adding CPU offloading and reducing it fitable to 8 GB gpus if be doable
Second is fixing FP16 issue so I can hopefully make a free Kaggle account notebook for people who can't afford GPU

And here 1 image that failed. My followers trying to upscale a movie and such images are problematic

[JasonGUTU]

This is a great model. It literally destroys the very expensive SaaS Magnific. I compared with 10 unique images here : #38

My first suggestion would be adding CPU offloading and reducing it fitable to 8 GB gpus if be doable Second is fixing FP16 issue so I can hopefully make a free Kaggle account notebook for people who can't afford GPU

And here 1 image that failed. My followers trying to upscale a movie and such images are problematic

Yes the motion and degradation in video is different to images. We plan to make a video restoration large model in the future. We are raising resources and planing for this.

[FurkanGozukara]

One another suggestion from people is that being able to change the used SDXL model on the fly. Currently all models are loaded from single yaml file.

[ivaxsirc]

It's awesome a next level to restore old resolution, as you see the image it's perfect but the face needs improve, with topaz ai restore the face and the result is incredible, so SUPIR will be the king with a FPGAN, Codeformer, adetailer, or roop
The first is original, the second SUPIR, the third with the results of SUPIR restore with Topaz Photo AI

[eriklgamedev]

The clothes are perfect, the background, but the faces it's the only thing than cannot do with good results

This. Faces get warped badly. Gonna have to run a face restoration after upscaling.

[ivaxsirc]

The clothes are perfect, the background, but the faces it's the only thing than cannot do with good results, in topaz ai, the faces good but the clothes it's not the same, nothing to do with SUPIR

3f-33479b824e37)

[Fanghua-Yu]

@ivaxsirc We have not applied face restoration pipeline yet. Generally, like CodeFormer, it requires three stage, 1) face detaction 2) restoring each face independently 3) pasting face back.
I'm working on it now. A face restoration script will be released in a few days.

[ivaxsirc]

@ivaxsirc We have not applied face restoration pipeline yet. Generally, like CodeFormer, it requires three stage, 1) face detaction 2) restoring each face independently 3) pasting face back. I'm working on it now. A face restoration script will be released in a few days.

Thanks for your answer, i think that SUPIR mark and before and after in the restoration, another suggestion in a older photos with scratches the restoration is not successful, maybe with before works https://github.com/sanje2v/OldPhotoRestorationDLST

thks for your excellent work

[ivaxsirc]

With scratches images, i try with stage 1 too but scratches persist

[tongchangD]

@ivaxsirc 老照片修复 项目

[ivaxsirc]

@ivaxsirc 老照片修复 项目

Thanks stable diffusion has a extension restore old photos that repair scratched images perfectly

[ivaxsirc]

It's Incredible, it's a shame the color, are there any way to repair color in a sun damage photo?

[Fanghua-Yu]

It's awesome a next level to restore old resolution, as you see the image it's perfect but the face needs improve, with topaz ai restore the face and the result is incredible, so SUPIR will be the king with a FPGAN, Codeformer, adetailer, or roop The first is original, the second SUPIR, the third with the results of SUPIR restore with Topaz Photo AI

@ivaxsirc Maybe you can try the new updated gradio_demo_face.py. It's an example of implementing SUPIR into the face restoration pipeline.
For this case, it restores face independently.


And paste back to the restored background.

[Fanghua-Yu]

It's Incredible, it's a shame the color, are there any way to repair color in a sun damage photo? @ivaxsirc

Currently, colors of restored images are strictly aligned with LQ. Repairing sun damage may require combining other projects.

[ivaxsirc]

It's awesome a next level to restore old resolution, as you see the image it's perfect but the face needs improve, with topaz ai restore the face and the result is incredible, so SUPIR will be the king with a FPGAN, Codeformer, adetailer, or roop The first is original, the second SUPIR, the third with the results of SUPIR restore with Topaz Photo AI

@ivaxsirc Maybe you can try the new updated gradio_demo_face.py. It's an example of implementing SUPIR into the face restoration pipeline. For this case, it restores face independently. And paste back to the restored background.

Yes, i can try, i have several older movies of 70's that we want see if it's possible restore with a 4k quality, and with SUPIR i think that will be possible, SUPIR could be a referent to restauration older movies, and why not in SVD

[ivaxsirc]

It's Incredible, it's a shame the color, are there any way to repair color in a sun damage photo? @ivaxsirc

Currently, colors of restored images are strictly aligned with LQ. Repairing sun damage may require combining other projects.

I know, at this moment SUPIR with the restoration of images plus the restoration of faces that you have shown me, is unbeatable, but if I were also able to fix old photos with scratches and with images of color worn by the sun, or the years, then I would be light years ahead of everyone, because there is no tool that can do that, only topaz with the color in beta, but the overall restoration quality is very, very below supir, even from what you have shown me with the faces.

I always make the proposals with a constructive purpose so that you are a reference in the subject of image restoration, which you already are, because I suppose that you have realized that with supir all the upscaler models are over, since it is superior to all of them.

The only problem is the graphics requirements, but if Forge will implement it I think that even with 8GB VRAM cards it could be used.

[barepixels]

The RAM requirement depends on how big the image you start with and how big you want to go

How does this work out... Fix the damaged photo with Topaz Photo AI first then SUPIR

[ivaxsirc]

Train the module of face restoration, awesome

[ivaxsirc]

I tried with extreme conditions
This is original image

This is Jennifer love Hewitt in a serie, if you put the name of actress with this parameters

prompt: a very young Jennifer love hewitt, eyes closed
a_prompt: Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore detailing, hyper sharpness, perfect without deformations.
n_prompt: painting, oil painting, illustration, drawing, art, sketch, oil painting, cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth
num_samples: 1
upscale: 1
edm_steps: 104
s_stage1: -1
s_stage2: 0.75
s_cfg: 1.7
seed: 174322976
s_churn: 1
s_noise: 1.003
color_fix_type: Wavelet
diff_dtype: fp16
ae_dtype: bf16
gamma_correction: 1
linear_CFG: True
linear_s_stage2: False
spt_linear_CFG: 7.5
spt_linear_s_stage2: 0
model_select: v0-F

If you put this other parameters
prompt: a very young girl, eyes closed
a_prompt: Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore detailing, hyper sharpness, perfect without deformations.
n_prompt: painting, oil painting, illustration, drawing, art, sketch, oil painting, cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth
num_samples: 1
upscale: 1
edm_steps: 104
s_stage1: -1
s_stage2: 0.75
s_cfg: 1.7
seed: 1736521820
s_churn: 1
s_noise: 1.003
color_fix_type: Wavelet
diff_dtype: fp16
ae_dtype: bf16
gamma_correction: 1
linear_CFG: True
linear_s_stage2: False
spt_linear_CFG: 7.5
spt_linear_s_stage2: 0
model_select: v0-F

therefore it interprets the keyword of the prompt, in this case Jennifer love Hewitt, but she looks older, much older than the other image where it only says very young girl, surely the reason is because the checkpoint is trained with the older actress and not when young. But this test also means that you can train your own checkpoint models and when you call them at the prompt they will appear.

It is like a faceswapper but depending on the highest quality trained dataset more better

But one important thing is also perceived (and it is not a problem of SUPIR, it is a problem of the checkpoints) is that it does not know how to interpret closed eyes, of all the tests carried out none have been done with eyes closed, which means that in the trained model checkpoint there are no faces with closed eyes.

I'm right?

[eriklgamedev]

I find adding "face wrinkles" to the negative prompt removes all the aging effects on faces.

I tried with extreme conditions This is original image This is Jennifer love Hewitt in a serie, if you put the name of actress with this parameters prompt: a very young Jennifer love hewitt, eyes closed a_prompt: Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore detailing, hyper sharpness, perfect without deformations. n_prompt: painting, oil painting, illustration, drawing, art, sketch, oil painting, cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth num_samples: 1 upscale: 1 edm_steps: 104 s_stage1: -1 s_stage2: 0.75 s_cfg: 1.7 seed: 174322976 s_churn: 1 s_noise: 1.003 color_fix_type: Wavelet diff_dtype: fp16 ae_dtype: bf16 gamma_correction: 1 linear_CFG: True linear_s_stage2: False spt_linear_CFG: 7.5 spt_linear_s_stage2: 0 model_select: v0-F

If you put this other parameters prompt: a very young girl, eyes closed a_prompt: Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore detailing, hyper sharpness, perfect without deformations. n_prompt: painting, oil painting, illustration, drawing, art, sketch, oil painting, cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth num_samples: 1 upscale: 1 edm_steps: 104 s_stage1: -1 s_stage2: 0.75 s_cfg: 1.7 seed: 1736521820 s_churn: 1 s_noise: 1.003 color_fix_type: Wavelet diff_dtype: fp16 ae_dtype: bf16 gamma_correction: 1 linear_CFG: True linear_s_stage2: False spt_linear_CFG: 7.5 spt_linear_s_stage2: 0 model_select: v0-F

therefore it interprets the keyword of the prompt, in this case Jennifer love Hewitt, but she looks older, much older than the other image where it only says very young girl, surely the reason is because the checkpoint is trained with the older actress and not when young. But this test also means that you can train your own checkpoint models and when you call them at the prompt they will appear.

It is like a faceswapper but depending on the highest quality trained dataset more better

But one important thing is also perceived (and it is not a problem of SUPIR, it is a problem of the checkpoints) is that it does not know how to interpret closed eyes, of all the tests carried out none have been done with eyes closed, which means that in the trained model checkpoint there are no faces with closed eyes.

I'm right?

[RexLeeGrey]

supir performs well most of the time, but for certain textures it's terrible






Stage1 options
▼
Gamma Correction
1
Stage2 options
▼
Number Of Images To Generate
1
Batch Size
1
Upscale
2
Randomize Seed
Steps
50
Text Guidance Scale
7.5
Stage2 Guidance Strength
1
Stage1 Guidance Strength
-1
Seed
2128147122
S-Churn
5
S-Noise
1.003
Default Positive Prompt
Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore detailing, hyper sharpness, perfect without deformations.
Default Negative Prompt
painting, oil painting, illustration, drawing, art, sketch, oil painting, cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth
Upscaled Images Output
Stage1 Run
LlaVa Run
Stage2 Run
Batch Processing Input Folder Path - If image_file_name.txt exists it will be read and used as prompt (optional). Uses same settings of single upscale (Stage 2 Run). If no caption txt it will use the Prompt you written. It can be empty as well.
e.g. /workspace/SUPIR_video/comparison_images
Batch Processing Output Folder Path - If left empty images are saved in default folder
outputs
Start Batch Upscaling
标签
Batch Processing Progress
Param Setting
Fidelity
Reset Param
Linear CFG
CFG Start
4
Linear Stage2 Guidance
Guidance Start
0

Diffusion Data Type
fp32
fp16
bf16
Auto-Encoder Data Type
fp32
bf16
Color-Fix Type
None
AdaIn
Wavelet
Model Selection
v0-Q
v0-F
LLaVA options
▼
Temperature
0.2
Top P
0.7

[ivaxsirc]

supir performs well most of the time, but for certain textures it's terrible Stage1 options ▼ Gamma Correction 1 Stage2 options ▼ Number Of Images To Generate 1 Batch Size 1 Upscale 2 Randomize Seed Steps 50 Text Guidance Scale 7.5 Stage2 Guidance Strength 1 Stage1 Guidance Strength -1 Seed 2128147122 S-Churn 5 S-Noise 1.003 Default Positive Prompt Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore detailing, hyper sharpness, perfect without deformations. Default Negative Prompt painting, oil painting, illustration, drawing, art, sketch, oil painting, cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth Upscaled Images Output Stage1 Run LlaVa Run Stage2 Run Batch Processing Input Folder Path - If image_file_name.txt exists it will be read and used as prompt (optional). Uses same settings of single upscale (Stage 2 Run). If no caption txt it will use the Prompt you written. It can be empty as well. e.g. /workspace/SUPIR_video/comparison_images Batch Processing Output Folder Path - If left empty images are saved in default folder outputs Start Batch Upscaling 标签 Batch Processing Progress Param Setting Fidelity Reset Param Linear CFG CFG Start 4 Linear Stage2 Guidance Guidance Start 0

Diffusion Data Type fp32 fp16 bf16 Auto-Encoder Data Type fp32 bf16 Color-Fix Type None AdaIn Wavelet Model Selection v0-Q v0-F LLaVA options ▼ Temperature 0.2 Top P 0.7

What prompt you use? With llava you can interprete the image.
SUPIR can do a thing very important, if you train your own dataset, you can create your own checkpoint, and maybe the textures will be perfect

[RexLeeGrey]

I'm used llava to generate the prompt

[FurkanGozukara]

we just added a new feature generate comparison video and working great

for Juggernaut-XL_v9 lower Text Guidance Scale working better. if set to 7.5, hallucination increases

if set to 6 works better

blade_0011.mp4

[ivaxsirc]

I'm used llava to generate the prompt

Could i try with your images?

[ivaxsirc]

The first video with restauration of SUPIR

Any idea for consistency image?

Video.de.WhatsApp.2024-03-05.a.las.23.13.52_fa8c53a2.mp4

[eriklgamedev]

The first video with restauration of SUPIR

Any idea for consistency image?

Video.de.WhatsApp.2024-03-05.a.las.23.13.52_fa8c53a2.mp4

Fix the seed seems to help with consistency between generations.

[RexLeeGrey]

I'm used llava to generate the prompt

Could i try with your images?

Yes, please help me try to fix it

[Fanghua-Yu]

The first video with restauration of SUPIR
Any idea for consistency image?
Video.de.WhatsApp.2024-03-05.a.las.23.13.52_fa8c53a2.mp4

Fix the seed seems to help with consistency between generations.

Using a fixed noise map can mitigate video consistency issues in a degree. However, high-frequency details will stick to the corresponding pixel, and will not move with objects.

For video restoration, tuning image model and adding temporal information are vital.
There is still a long way to go for a video restoration model with high fidelity, quality and consistency. Currently, we are collecting data for it :)

[ivaxsirc]

supir performs well most of the time, but for certain textures it's terrible Stage1 options ▼ Gamma Correction 1 Stage2 options ▼ Number Of Images To Generate 1 Batch Size 1 Upscale 2 Randomize Seed Steps 50 Text Guidance Scale 7.5 Stage2 Guidance Strength 1 Stage1 Guidance Strength -1 Seed 2128147122 S-Churn 5 S-Noise 1.003 Default Positive Prompt Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore detailing, hyper sharpness, perfect without deformations. Default Negative Prompt painting, oil painting, illustration, drawing, art, sketch, oil painting, cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth Upscaled Images Output Stage1 Run LlaVa Run Stage2 Run Batch Processing Input Folder Path - If image_file_name.txt exists it will be read and used as prompt (optional). Uses same settings of single upscale (Stage 2 Run). If no caption txt it will use the Prompt you written. It can be empty as well. e.g. /workspace/SUPIR_video/comparison_images Batch Processing Output Folder Path - If left empty images are saved in default folder outputs Start Batch Upscaling 标签 Batch Processing Progress Param Setting Fidelity Reset Param Linear CFG CFG Start 4 Linear Stage2 Guidance Guidance Start 0

Diffusion Data Type fp32 fp16 bf16 Auto-Encoder Data Type fp32 bf16 Color-Fix Type None AdaIn Wavelet Model Selection v0-Q v0-F LLaVA options ▼ Temperature 0.2 Top P 0.7

Is wood or stone?

[RexLeeGrey]

supir performs well most of the time, but for certain textures it's terrible Stage1 options ▼ Gamma Correction 1 Stage2 options ▼ Number Of Images To Generate 1 Batch Size 1 Upscale 2 Randomize Seed Steps 50 Text Guidance Scale 7.5 Stage2 Guidance Strength 1 Stage1 Guidance Strength -1 Seed 2128147122 S-Churn 5 S-Noise 1.003 Default Positive Prompt Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore detailing, hyper sharpness, perfect without deformations. Default Negative Prompt painting, oil painting, illustration, drawing, art, sketch, oil painting, cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth Upscaled Images Output Stage1 Run LlaVa Run Stage2 Run Batch Processing Input Folder Path - If image_file_name.txt exists it will be read and used as prompt (optional). Uses same settings of single upscale (Stage 2 Run). If no caption txt it will use the Prompt you written. It can be empty as well. e.g. /workspace/SUPIR_video/comparison_images Batch Processing Output Folder Path - If left empty images are saved in default folder outputs Start Batch Upscaling 标签 Batch Processing Progress Param Setting Fidelity Reset Param Linear CFG CFG Start 4 Linear Stage2 Guidance Guidance Start 0大部分情况下，supir 表现良好，但对于某些纹理来说，效果很糟糕 < b5> 第 1 阶段选项 ▼ 伽玛校正 1 第 2 阶段选项 ▼ 生成的图像数量 1 批量大小 1 放大 2 随机种子步骤 50 文本指导​​比例 7.5 第 2 阶段指导强度 1 第 1 阶段指导强度 -1 种子 2128147122 S-Churn 5 S - 噪点 1.003 默认正提示电影，高对比度，高度详细，使用佳能 EOS R 相机拍摄，超详细照片 - 逼真的最大细节，32k，颜色分级，超高清，极其细致的细节，皮肤毛孔细节，超锐度，完美无变形。默认负提示绘画、油画、插图、绘画、艺术、素描、油画、卡通、CG风格、3D渲染、虚幻引擎、模糊、肮脏、凌乱、最差质量、低质量、框架、水印、签名、jpeg伪像、变形、低分辨率、过度平滑放大图像输出阶段 1 运行 LlaVa 运行阶段 2 运行批处理输入文件夹路径 - 如果 image_file_name.txt 存在，它将被读取并用作提示（可选）。使用与单一高档（第 2 阶段运行）相同的设置。如果没有标题 txt，它将使用您编写的提示。它也可以为空。例如/workspace/SUPIR_video/comparison_images 批处理输出文件夹路径 - 如果留空图像保存在默认文件夹输出中 开始批量放大 标签 批处理进度参数设置保真度重置参数 Linear CFG CFG Start 4 Linear Stage2 Guidance Guidance Start 0
Diffusion Data Type fp32 fp16 bf16 Auto-Encoder Data Type fp32 bf16 Color-Fix Type None AdaIn Wavelet Model Selection v0-Q v0-F LLaVA options ▼ Temperature 0.2 Top P 0.7扩散数据类型 fp32 fp16 bf16 自动编码器数据类型 fp32 bf16 颜色固定类型 无 AdaIn 小波模型选择 v0-Q v0-F LLaVA 选项 ▼ 温度 0.2 Top P 0.7

Is wood or stone? 重试    错误原因

It's wood carving, wood chiseling. Wood carving texture.

[ivaxsirc]

More tested on face restoration

Original image

Image supir without checked anything

The image is good, but the face is not similar to the original

Same parameters face restoration checked

The face is restored but it's not similar to the woman original

Same parameters BG and face restoration checked

The global quality is very good but the face is not similar to the original

I used Forge and 2 controlnets 1) mediapipe_face

2. FaceSwaplab with a reference image

With inpaint result

The face image does not have much quality

But with SUPIR...

Comparison

This is the face of the real artist

As you can see there is a great similarity in the faces, now all you have to do is join the face with the part of the image already restored, it would be great if SUPIR had some tool to do it automatically

In conclusion, with face images of very low resolution, SUPIR cannot save the fidelity of similarity, but if it were able to understand a reference image as you have seen, the similarity would be very high

[Feynman1999]

The first video with restauration of SUPIR
Any idea for consistency image?
Video.de.WhatsApp.2024-03-05.a.las.23.13.52_fa8c53a2.mp4

Fix the seed seems to help with consistency between generations.

Using a fixed noise map can mitigate video consistency issues in a degree. However, high-frequency details will stick to the corresponding pixel, and will not move with objects.

For video restoration, tuning image model and adding temporal information are vital. There is still a long way to go for a video restoration model with high fidelity, quality and consistency. Currently, we are collecting data for it :)

The first video with restauration of SUPIR
Any idea for consistency image?
Video.de.WhatsApp.2024-03-05.a.las.23.13.52_fa8c53a2.mp4

Fix the seed seems to help with consistency between generations.

Using a fixed noise map can mitigate video consistency issues in a degree. However, high-frequency details will stick to the corresponding pixel, and will not move with objects.

For video restoration, tuning image model and adding temporal information are vital. There is still a long way to go for a video restoration model with high fidelity, quality and consistency. Currently, we are collecting data for it :)

The first video with restauration of SUPIR
Any idea for consistency image?
Video.de.WhatsApp.2024-03-05.a.las.23.13.52_fa8c53a2.mp4

Fix the seed seems to help with consistency between generations.

Using a fixed noise map can mitigate video consistency issues in a degree. However, high-frequency details will stick to the corresponding pixel, and will not move with objects.

For video restoration, tuning image model and adding temporal information are vital. There is still a long way to go for a video restoration model with high fidelity, quality and consistency. Currently, we are collecting data for it :)

If the noise level is lowered through the parameter s_noise, how does this value affect image noise? Can you explain Why is the default value a 1.003

[lschaupp]

Can anyone (who has a bit of time) try this here for animation: https://www.reddit.com/r/StableDiffusion/comments/1bk9e0d/smooth_diffusion_code_released/ in combination with SUPIR? Could potentially lead to less artifacts when using SUPIR for a video.

[CuddleSabe]

Can anyone (who has a bit of time) try this here for animation: https://www.reddit.com/r/StableDiffusion/comments/1bk9e0d/smooth_diffusion_code_released/ in combination with SUPIR? Could potentially lead to less artifacts when using SUPIR for a video.

I tried it in last month, but I found my image quality has been worse.

[huanghaosen110]

我对一张VGGface的图片进行高清化，但生成的人脸图像油光满面，眼睛变得不清晰，请问是什么问题？

我仅仅将min_size 设置为512，no_llava=True,其他的都是默认配置

[JasonGUTU]

我对一张VGGface的图片进行高清化，但生成的人脸图像油光满面，眼睛变得不清晰，请问是什么问题？

我仅仅将min_size 设置为512，no_llava=True,其他的都是默认配置

Please share your input with us.

[CuddleSabe]

我对一张VGGface的图片进行高清化，但生成的人脸图像油光满面，眼睛变得不清晰，请问是什么问题？

我仅仅将min_size 设置为512，no_llava=True,其他的都是默认配置

因为你跑图尺寸用的512，但是编码original_sizhe和target_size还是1024，而sdxl跑图尺寸512会有问题。