diff --git a/docs.json b/docs.json index 01bedacf8..71d1cbcfb 100644 --- a/docs.json +++ b/docs.json @@ -266,7 +266,9 @@ "group": "Utility", "pages": [ "tutorials/utility/preprocessors", - "tutorials/utility/frame-interpolation" + "tutorials/utility/frame-interpolation", + "tutorials/utility/image-upscale", + "tutorials/utility/video-upscale" ] }, { @@ -989,7 +991,9 @@ "group": "Utility", "pages": [ "zh-CN/tutorials/utility/preprocessors", - "zh-CN/tutorials/utility/frame-interpolation" + "zh-CN/tutorials/utility/frame-interpolation", + "zh-CN/tutorials/utility/image-upscale", + "zh-CN/tutorials/utility/video-upscale" ] }, { diff --git a/tutorials/basic/upscale.mdx b/tutorials/basic/upscale.mdx index b96ca04d0..4a93da2e4 100644 --- a/tutorials/basic/upscale.mdx +++ b/tutorials/basic/upscale.mdx @@ -17,6 +17,10 @@ This article covers one of many upscaling methods in ComfyUI. In this tutorial, 2. Performing basic image upscaling 3. Combining text-to-image workflows with upscaling + +Looking for more advanced upscaling options? Check out our comprehensive [Image Upscaling Guide](/tutorials/utility/image-upscale) which covers various models and use cases including portrait enhancement, product photography, and more. + + ## Upscaling Workflow ### Model Installation @@ -87,4 +91,4 @@ Model characteristics: 1. **Chained Upscaling**: Combine multiple upscale nodes (e.g., 2x → 4x) for ultra-high magnification 2. **Hybrid Workflow**: Connect upscale nodes after generation for "generate+enhance" pipelines -3. **Comparative Testing**: Different models perform better on specific image types - test multiple options \ No newline at end of file +3. **Comparative Testing**: Different models perform better on specific image types - test multiple options diff --git a/tutorials/utility/image-upscale.mdx b/tutorials/utility/image-upscale.mdx new file mode 100644 index 000000000..cad2c3271 --- /dev/null +++ b/tutorials/utility/image-upscale.mdx @@ -0,0 +1,192 @@ +--- +title: "Image upscaling in ComfyUI" +description: "Learn how to upscale images in ComfyUI using local models and partner nodes for different use cases" +sidebarTitle: "Image upscale" +--- + +This guide covers image upscaling workflows in ComfyUI, including local models and partner node options for various use cases. + + + For a comprehensive deep-dive with benchmarks and visual comparisons, read the full article on the ComfyUI Blog. + + +## Why upscaling matters + +- **Industry requirements for resolutions**: 4K/8K end results are widely required across VFX & Film, Marketing, E-commerce, Gaming, and Design industries, and requirements keep moving up. +- **Reframing**: When assets ship across different aspect ratios and placements, extra resolution headroom prevents quality loss after cropping or recomposition. +- **The AI content problem**: Generative AI has dramatically increased the speed of visual creation, but most AI assets today are still generated at 480–720p, and not production-ready at the pixel level: edges, micro-texture, faces/hands, compression artifacts, etc. This creates a new pipeline: Generation → Refinement/Fix → Upscale → Delivery. +- **Cost/time efficiency**: Generate or render smaller → upscale at the end. This is a standard efficiency move in pipelines under budget/time pressure. + +## Key concepts + +### Upscaling vs. enhancement + +- **Upscaling** increases resolution while reconstructing detail. +- **Enhancement** improves perceived quality (denoise, sharpen, restore, color, faces, etc.). + +The best practice in AI pipelines usually combines both steps. + +### Creative vs. conservative upscale + +Diffusion models and generative AI changed what "upscaling" means. Traditionally, super-resolution aimed to **preserve the original signal**. Today, some models can **re-imagine details that never existed**. And there are models approaching different balances between the two. + +| | Conservative | Creative | +|---|---|---| +| **Approach** | Preserve the original | Reimagine and enhance | +| **Strengths** | High accuracy, consistency, production-safe | Rich details, sharper, more visually striking | +| **Limits** | Can appear flat, limited improvement on poor input | Possible hallucinations or structural drift | +| **Models** | Magnific Precise, SeedVR2, FlashVSR, Topaz Fast, HitPaw | Wan 2.2, Magnific Creative, Topaz Astra, HitPaw Creative | + +## Use cases + + +**TL;DR** +- **Portraits**: Magnific Skin Enhancer +- **Product photography**: Magnific Precise, WaveSpeed SeedVR2, or Nano Banana Pro +- **Landscapes & illustrations**: Model choice depends on your specific needs +- **AI artifacts**: Do not rely on upscaling to fix common AI artifacts +- **SeedVR2 tip**: Downscale the image to 0.35 megapixels with the `ImageScaleToTotalPixels` node before upscaling for better results + + +### Portrait and skin enhancement + +When it comes to upscaling portraits of realistic subjects, achieving realistic skin details while maintaining character consistency is key. The best upscale models for fixing plastic skin need to add texture, pores and natural skin imperfections. In this domain, Magnific Image Skin Enhancer far outperforms the rest. + +**Recommended**: Magnific Image Skin Enhancer + + + + Try on Comfy Cloud or download workflow + + + +### Product photography + +Maintaining a faithful representation of materials, product label edges, and small text is a must when upscaling product images. Therefore, conservative upscaling models are necessary. + +**Recommended**: HitPaw, Magnific Precise, WaveSpeed SeedVR2, Nano Banana Pro + + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + +### Landscape and environments + +For this use case, the upscale model you select really depends on your needs. Have an environmental shot that you want to pop with details and set a mood? Use a creative upscale model. Have an establishing shot where a building needs to stay consistent? Use a conservative upscale model. + +It's worth noting that if your input image has artifacts, creative upscale models may be able to reimagine the artifact while conservative upscale models will not. + + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + +### Stylized art and illustration + +This is another use case where the correct upscale model depends on your needs. A rule of thumb is that if your input has a very distinct/unique style, the best model will be conservative. Creative models may add too many details and stray away from the desired illustrated style. However, if your input image has room for detail, a creative model works well and can even improve the style. + +We recommend playing around with the 'creativity' parameters for Magnific Creative and Topaz Image Enhance to find values that meet your needs! Nano Banana Pro can also work well for more common styles and adding details (but may rely on 'seed luck'). + +**Recommended**: Magnific Creative, Topaz Image Enhance, Nano Banana Pro + + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + +### AI-generated images + +When it comes to upscaling images that have typical AI issues such as too many fingers, artifacts, incorrect anatomy, and morphing, the assumption is that creative upscales can fix it. Sometimes that's true, but other times not so much. Best practice is to resolve these issues prior to upscaling with image-edit models or traditional tools (or generate a new image entirely). + +## Available models + +### Local models (ESRGAN) + +For basic local upscaling using ESRGAN models, see the [basic upscale tutorial](/tutorials/basic/upscale). + +| Model | Best for | +|---|---| +| RealESRGAN | General-purpose upscaling | +| BSRGAN | Text and sharp edges | +| SwinIR | Natural textures, landscapes | + + + + Try on Comfy Cloud or download workflow + + + +### General creative upscaling + +For general creative upscaling across various use cases: + + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + +### Partner nodes + +Partner nodes provide access to advanced upscaling models via API. + +| Model | Type | Features | +|---|---|---| +| Topaz Image Enhance (Bloom) | Creative | Subject detection, face enhancement, color preservation, up to 8K | +| Magnific Precise | Conservative | High accuracy, production-safe | +| Magnific Creative | Creative | Detail reimagination | +| Magnific Skin Enhancer | Creative | Portrait-specific, adds realistic skin texture | +| Nano Banana Pro | Conservative | Fast, good for product photography | +| WaveSpeed SeedVR2 | Conservative | High fidelity | +| HitPaw | Both | Conservative and creative modes | +| Recraft | Both | Creative and crisp modes | + +## Benchmark: 1K to 4K upscale time + +| Model | Time | +|---|---| +| Magnific Precise | ~40s | +| WaveSpeed SeedVR2 | ~40s | +| Magnific Creative | ~50s | +| Magnific Skin Enhancer | ~60s | +| HitPaw | ~60s | +| Nano Banana Pro | ~80s | +| Topaz Image Enhance | ~100s | + +## Tips + +- For SeedVR2, downscale the image to 0.35 megapixels with the `ImageScaleToTotalPixels` node before upscaling for better results. +- Chain multiple upscale nodes (e.g., 2x → 4x) for ultra-high magnification. +- Connect upscale nodes after generation for "generate + enhance" pipelines. +- Test multiple models on your specific image type to find the best fit. diff --git a/tutorials/utility/video-upscale.mdx b/tutorials/utility/video-upscale.mdx new file mode 100644 index 000000000..a929fbfee --- /dev/null +++ b/tutorials/utility/video-upscale.mdx @@ -0,0 +1,184 @@ +--- +title: "Video upscaling in ComfyUI" +description: "Learn how to upscale videos in ComfyUI using local models and partner nodes for different use cases" +sidebarTitle: "Video upscale" +--- + +This guide covers video upscaling workflows in ComfyUI, including local models and partner node options for various use cases. + + + For a comprehensive deep-dive with benchmarks and visual comparisons, see the full article on the ComfyUI Blog. + + +## TL;DR + +- **Quality mode**: SeedVR2 and HitPaw deliver the most accurate and consistent realism. +- **Speed mode**: FlashVSR and HitPaw are the fastest options. +- **AI video fix**: Topaz Astra Creative is strongest for cinematic polish and fixing AI-generated video. +- **Archival restoration**: SeedVR2 preserves original structure while enhancing clarity. +- **CGI/3D render**: Topaz, SeedVR2, and HitPaw all produce clean, sharp results. + +## Key concepts + +### Upscaling vs. enhancement + +- **Upscaling** increases resolution while reconstructing detail. +- **Enhancement** improves perceived quality (denoise, sharpen, restore, temporal consistency, etc.). + +Video upscaling adds temporal considerations—maintaining consistency across frames while improving resolution. + +### Creative vs. conservative upscale + +| | Conservative | Creative | +|---|---|---| +| **Approach** | Preserve the original | Reimagine and enhance | +| **Strengths** | High accuracy, temporal consistency, production-safe | Rich details, sharper, more visually striking | +| **Limits** | Can appear flat, limited improvement on poor input | Possible hallucinations, temporal flickering | +| **Models** | Topaz Starlight Fast, FlashVSR, SeedVR2, HitPaw | Topaz Starlight Creative (Astra), Wan 2.2, HitPaw | + +## Use cases + +### Quality mode + +**SeedVR2** and **HitPaw** deliver the most accurate results, with strong image quality and excellent character consistency. + +**Topaz** also upscales effectively, but it can sometimes be more creative than expected. This makes it especially suitable for cinematic footage where a stylized look is acceptable. + + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + +### Speed mode + +**FlashVSR** performs extremely well when speed is the priority. Upscaling from 720p to 1080p takes about **41 seconds for a 10-second video**, making it one of the fastest options available. + +**HitPaw** is also relatively fast. In **Topaz**, the Fast and Creative modes render at similar speeds, so using **Creative** mode is generally recommended for better visual results without sacrificing time. + + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + +### AI video fix + +**Topaz Astra Creative** produces very smooth and visually consistent outputs, making it a strong choice for repairing AI-generated videos. + +A useful workflow tip is to first run a **fix or cleanup pass**, then upscale to **1080p**, and finally upscale gradually to **4K**. This step-by-step approach usually leads to better stability and detail. + + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + +### Archival restoration + +**SeedVR2** is particularly impressive for archival restoration. It preserves the original structure and detail of the source footage while enhancing clarity, avoiding the introduction of artificial or exaggerated features. + + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + +### CGI / 3D render + +**Topaz, SeedVR2 and HitPaw** all perform very well on CGI and 3D-rendered footage. They produce clean, sharp results and maintain the crisp edges that are important for synthetic imagery. + + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + Try on Comfy Cloud or download workflow + + + +### Animation and motion graphics + +Clean lines and consistent colors are essential. Conservative upscaling preserves the intended style. + +**Recommended**: Topaz Starlight Fast, Fast GAN + + + + Try on Comfy Cloud or download workflow + + + +## Upscale speed comparison (10s 720p clips) + +| Model | Resolution | Time | +|---|---|---| +| FlashVSR (local) | 1080p | ~41s | +| FlashVSR (local) | 4K | ~52s | +| HitPaw | 2K | ~80s | +| HitPaw | 4K | ~175s | +| SeedVR2 (local, 5090) | 1080p | ~312s | +| Topaz Astra | 1080p | ~374s | +| Topaz Astra | 4K | ~560s | + +## Available models + +### Partner nodes + +Partner nodes provide access to advanced video upscaling models via API. + +| Model | Type | Max resolution | Features | +|---|---|---|---| +| Topaz Starlight Fast | Conservative | 4K | Fast, precision upscale | +| Topaz Starlight Creative (Astra) | Creative | 4K | Diffusion-based creative upscale | +| Topaz Apollo | - | - | Frame interpolation | +| WaveSpeed SeedVR2 | Conservative | - | High fidelity | +| WaveSpeed FlashVSR | Conservative | - | Fast video super-resolution | +| HitPaw | Both | - | All-in-one video enhancement | + +### Topaz Video Enhance features + +- **Upscaling**: Starlight (Astra) supports fast video upscaling to FullHD (1080p) or 4K while maintaining clarity. +- **Interpolation**: Apollo model for frame interpolation with slowmo settings and custom frame rates (e.g., 60fps). +- **Dynamic compression**: Intelligently identifies and handles compression artifacts to improve image purity. + +## Getting started with Topaz Video Enhance + +1. Update ComfyUI to the latest version. +2. Search for the `Topaz Video Enhance` node in the node list. +3. Connect your video input and configure the settings. +4. Run the workflow. + +## Tips + +- Combine upscaling with [frame interpolation](/tutorials/utility/frame-interpolation) for smoother, higher-resolution output. +- For AI-generated videos, consider fixing artifacts before upscaling rather than relying on creative upscaling to fix them. +- Test different models on sample clips before processing full videos. +- Use conservative upscaling for content that requires strict fidelity to the original. diff --git a/zh-CN/tutorials/basic/upscale.mdx b/zh-CN/tutorials/basic/upscale.mdx index 11812fcce..e56f53bc7 100644 --- a/zh-CN/tutorials/basic/upscale.mdx +++ b/zh-CN/tutorials/basic/upscale.mdx @@ -16,6 +16,9 @@ import InstallationModels from '/snippets/zh/tutorials/basic/installation-models - 使用放大模型进行一次简单的放大 - 结合文生图工作流，完成图像的放大 + +想了解更多高级放大选项？请查看我们的综合[图像放大指南](/zh-CN/tutorials/utility/image-upscale)，涵盖各种模型和使用场景，包括人像增强、产品摄影等。 + ## 下载并安装放大模型 @@ -80,4 +83,4 @@ import InstallationModels from '/snippets/zh/tutorials/basic/installation-models 1. **链式放大**：对于需要超高倍率放大的情况，可以串联多个放大节点（如先2x再4x） 2. **混合放大**：在生成工作流后接放大节点，实现"生成+增强"一体化流程 -3. **对比测试**：不同模型对特定类型图片效果差异较大，建议同时测试多个模型 \ No newline at end of file +3. **对比测试**：不同模型对特定类型图片效果差异较大，建议同时测试多个模型 diff --git a/zh-CN/tutorials/utility/image-upscale.mdx b/zh-CN/tutorials/utility/image-upscale.mdx new file mode 100644 index 000000000..fedbd32ba --- /dev/null +++ b/zh-CN/tutorials/utility/image-upscale.mdx @@ -0,0 +1,192 @@ +--- +title: "ComfyUI 图像放大" +description: "学习如何在 ComfyUI 中使用本地模型和合作伙伴节点进行图像放大" +sidebarTitle: "图像放大" +--- + +本指南介绍 ComfyUI 中的图像放大工作流，包括本地模型和合作伙伴节点选项，适用于各种使用场景。 + + + 如需更全面的深入分析、基准测试和视觉对比，请阅读 ComfyUI 博客上的完整文章。 + + +## 为什么需要图像放大 + +- **行业对分辨率的要求**：4K/8K 最终成品在影视特效、营销、电商、游戏和设计行业中被广泛要求，且标准还在不断提高。 +- **重新构图**：当素材需要适配不同的宽高比和位置时，额外的分辨率余量可以防止裁剪或重新构图后的质量损失。 +- **AI 内容的问题**：生成式 AI 大幅提升了视觉创作的速度，但目前大多数 AI 素材仍然以 480-720p 生成，在像素级别上还未达到生产标准：边缘、微纹理、面部/手部、压缩伪影等。这催生了新的工作流程：生成 → 修复/优化 → 放大 → 交付。 +- **成本/时间优化**：先生成或渲染较小尺寸 → 最后再放大。这是预算/时间紧张的工作流程中的标准效率策略。 + +## 核心概念 + +### 放大 vs. 增强 + +- **放大（Upscaling）**：在重建细节的同时提高分辨率。 +- **增强（Enhancement）**：改善感知质量（降噪、锐化、修复、色彩、面部等）。 + +AI 工作流的最佳实践通常是将两个步骤结合使用。 + +### 创意放大 vs. 保守放大 + +扩散模型和生成式 AI 改变了"放大"的含义。传统上，超分辨率旨在**保留原始信号**。如今，一些模型可以**重新想象从未存在过的细节**。不同的模型在两者之间取得不同的平衡。 + +| | 保守放大 | 创意放大 | +|---|---|---| +| **方式** | 保留原始内容 | 重新想象并增强 | +| **优势** | 高精度、一致性、适合生产环境 | 细节丰富、更锐利、视觉效果更突出 | +| **局限** | 可能显得平淡，对低质量输入改善有限 | 可能产生幻觉或结构偏移 | +| **模型** | Magnific Precise、SeedVR2、FlashVSR、Topaz Fast、HitPaw | Wan 2.2、Magnific Creative、Topaz Astra、HitPaw Creative | + +## 使用场景 + + +**总结** +- **人像**：Magnific Skin Enhancer +- **产品摄影**：Magnific Precise、WaveSpeed SeedVR2 或 Nano Banana Pro +- **风景和插画**：模型选择取决于你的具体需求 +- **AI 伪影**：不要依赖放大来修复常见的 AI 伪影 +- **SeedVR2 技巧**：放大前先使用 `ImageScaleToTotalPixels` 节点将图像缩小到 0.35 百万像素，效果更好 + + +### 人像和皮肤增强 + +在放大真实人物的肖像时，实现逼真的皮肤细节同时保持角色一致性是关键。最好的放大模型需要能够添加纹理、毛孔和自然的皮肤瑕疵来修复塑料感的皮肤。在这个领域，Magnific Image Skin Enhancer 远超其他模型。 + +**推荐**：Magnific Image Skin Enhancer + + + + 在 Comfy Cloud 上试用或下载工作流 + + + +### 产品摄影 + +在放大产品图像时，必须保持材质、产品标签边缘和小文字的忠实呈现。因此需要使用保守型放大模型。 + +**推荐**：HitPaw、Magnific Precise、WaveSpeed SeedVR2、Nano Banana Pro + + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + +### 风景和环境 + +对于这个使用场景，选择哪个放大模型取决于你的需求。想让环境照片细节丰富并营造氛围？使用创意型放大模型。有一个需要保持建筑一致性的场景？使用保守型放大模型。 + +值得注意的是，如果你的输入图像有伪影，创意型放大模型可能能够重新想象这些伪影，而保守型放大模型则不会。 + + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + +### 风格化艺术/插画 + +这是另一个正确的放大模型取决于你需求的使用场景。经验法则是，如果你的输入有非常独特的风格，最好的模型是保守型的。创意型模型可能会添加过多细节，偏离所需的插画风格。然而，如果你的输入图像有细节提升的空间，创意型模型效果很好，甚至可以改善风格。 + +我们建议尝试调整 Magnific Creative 和 Topaz Image Enhance 的"创意度"参数来找到满足你需求的值！Nano Banana Pro 对于更常见的风格和添加细节也很有效（但可能依赖"种子运气"）。 + +**推荐**：Magnific Creative、Topaz Image Enhance、Nano Banana Pro + + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + +### AI 生成图像 + +当放大具有典型 AI 问题（如多余的手指、伪影、不正确的解剖结构和变形）的图像时，人们通常假设创意放大可以修复这些问题。有时确实如此，但很多时候并非如此。最佳做法是在放大之前使用图像编辑模型或传统工具解决这些问题（或重新生成图像）。 + +## 可用模型 + +### 本地模型（ESRGAN） + +有关使用 ESRGAN 模型进行基础本地放大，请参阅[基础放大教程](/zh-CN/tutorials/basic/upscale)。 + +| 模型 | 最适合 | +|---|---| +| RealESRGAN | 通用放大 | +| BSRGAN | 文字和锐利边缘 | +| SwinIR | 自然纹理、风景 | + + + + 在 Comfy Cloud 上试用或下载工作流 + + + +### 通用创意放大 + +适用于各种使用场景的通用创意放大： + + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + +### 合作伙伴节点 + +合作伙伴节点通过 API 提供高级放大模型访问。 + +| 模型 | 类型 | 特性 | +|---|---|---| +| Topaz Image Enhance (Bloom) | 创意 | 主体检测、面部增强、色彩保留，最高 8K | +| Magnific Precise | 保守 | 高精度、适合生产环境 | +| Magnific Creative | 创意 | 细节重新想象 | +| Magnific Skin Enhancer | 创意 | 人像专用，添加逼真皮肤纹理 | +| Nano Banana Pro | 保守 | 快速，适合产品摄影 | +| WaveSpeed SeedVR2 | 保守 | 高保真度 | +| HitPaw | 两者 | 保守和创意模式 | +| Recraft | 两者 | 创意和清晰模式 | + +## 基准测试：1K 到 4K 放大时间 + +| 模型 | 时间 | +|---|---| +| Magnific Precise | ~40秒 | +| WaveSpeed SeedVR2 | ~40秒 | +| Magnific Creative | ~50秒 | +| Magnific Skin Enhancer | ~60秒 | +| HitPaw | ~60秒 | +| Nano Banana Pro | ~80秒 | +| Topaz Image Enhance | ~100秒 | + +## 技巧 + +- 对于 SeedVR2，在放大之前使用 `ImageScaleToTotalPixels` 节点将图像缩小到 0.35 百万像素可获得更好的效果。 +- 链接多个放大节点（例如 2x → 4x）以实现超高倍率放大。 +- 在生成后连接放大节点，实现"生成 + 增强"工作流。 +- 在特定图像类型上测试多个模型以找到最佳匹配。 diff --git a/zh-CN/tutorials/utility/video-upscale.mdx b/zh-CN/tutorials/utility/video-upscale.mdx new file mode 100644 index 000000000..292153cec --- /dev/null +++ b/zh-CN/tutorials/utility/video-upscale.mdx @@ -0,0 +1,184 @@ +--- +title: "ComfyUI 视频放大" +description: "学习如何在 ComfyUI 中使用本地模型和合作伙伴节点进行视频放大" +sidebarTitle: "视频放大" +--- + +本指南介绍 ComfyUI 中的视频放大工作流，包括本地模型和合作伙伴节点选项，适用于各种使用场景。 + + + 如需更全面的深入分析、基准测试和视觉对比，请参阅 ComfyUI 博客上的完整文章。 + + +## TL;DR + +- **质量模式**：SeedVR2 和 HitPaw 提供最准确和一致的真实感效果。 +- **速度模式**：FlashVSR 和 HitPaw 是最快的选择。 +- **AI 视频修复**：Topaz Astra Creative 在电影级润色和修复 AI 生成视频方面表现最强。 +- **档案修复**：SeedVR2 在增强清晰度的同时保留原始结构。 +- **CGI/3D 渲染**：Topaz、SeedVR2 和 HitPaw 都能产生干净、锐利的结果。 + +## 核心概念 + +### 放大 vs. 增强 + +- **放大（Upscaling）**：在重建细节的同时提高分辨率。 +- **增强（Enhancement）**：改善感知质量（降噪、锐化、修复、时间一致性等）。 + +视频放大增加了时间维度的考量——在提高分辨率的同时保持帧间一致性。 + +### 创意放大 vs. 保守放大 + +| | 保守放大 | 创意放大 | +|---|---|---| +| **方式** | 保留原始内容 | 重新想象并增强 | +| **优势** | 高精度、时间一致性、适合生产环境 | 细节丰富、更锐利、视觉效果更突出 | +| **局限** | 可能显得平淡，对低质量输入改善有限 | 可能产生幻觉、时间闪烁 | +| **模型** | Topaz Starlight Fast、FlashVSR、SeedVR2、HitPaw | Topaz Starlight Creative (Astra)、Wan 2.2、HitPaw | + +## 使用场景 + +### 质量模式 + +**SeedVR2** 和 **HitPaw** 提供最准确的结果，具有出色的图像质量和优秀的角色一致性。 + +**Topaz** 也能有效放大，但有时可能比预期更具创意性。这使其特别适合可以接受风格化外观的电影素材。 + + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + +### 速度模式 + +**FlashVSR** 在速度优先时表现极佳。将 720p 放大到 1080p 大约需要 **41 秒处理 10 秒视频**，是目前最快的选项之一。 + +**HitPaw** 也相对较快。在 **Topaz** 中，Fast 和 Creative 模式的渲染速度相似，因此通常建议使用 **Creative** 模式以获得更好的视觉效果而不牺牲时间。 + + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + +### AI 视频修复 + +**Topaz Astra Creative** 产生非常平滑和视觉一致的输出，是修复 AI 生成视频的强力选择。 + +一个有用的工作流技巧是先运行 **修复或清理步骤**，然后放大到 **1080p**，最后逐步放大到 **4K**。这种分步方法通常能带来更好的稳定性和细节。 + + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + +### 档案修复 + +**SeedVR2** 在档案修复方面表现尤为出色。它在增强清晰度的同时保留源素材的原始结构和细节，避免引入人工或夸张的特征。 + + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + +### CGI / 3D 渲染 + +**Topaz、SeedVR2 和 HitPaw** 在 CGI 和 3D 渲染素材上都表现出色。它们产生干净、锐利的结果，并保持对合成图像至关重要的清晰边缘。 + + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + 在 Comfy Cloud 上试用或下载工作流 + + + +### 动画和动态图形 + +清晰的线条和一致的颜色至关重要。保守放大可以保留预期的风格。 + +**推荐**：Topaz Starlight Fast、Fast GAN + + + + 在 Comfy Cloud 上试用或下载工作流 + + + +## 放大速度对比（10 秒 720p 视频） + +| 模型 | 分辨率 | 时间 | +|---|---|---| +| FlashVSR（本地） | 1080p | ~41s | +| FlashVSR（本地） | 4K | ~52s | +| HitPaw | 2K | ~80s | +| HitPaw | 4K | ~175s | +| SeedVR2（本地，5090） | 1080p | ~312s | +| Topaz Astra | 1080p | ~374s | +| Topaz Astra | 4K | ~560s | + +## 可用模型 + +### 合作伙伴节点 + +合作伙伴节点通过 API 提供高级视频放大模型访问。 + +| 模型 | 类型 | 最大分辨率 | 特性 | +|---|---|---|---| +| Topaz Starlight Fast | 保守 | 4K | 快速、精确放大 | +| Topaz Starlight Creative (Astra) | 创意 | 4K | 基于扩散的创意放大 | +| Topaz Apollo | - | - | 帧插值 | +| WaveSpeed SeedVR2 | 保守 | - | 高保真度 | +| WaveSpeed FlashVSR | 保守 | - | 快速视频超分辨率 | +| HitPaw | 两者 | - | 一体化视频增强 | + +### Topaz Video Enhance 功能 + +- **放大**：Starlight (Astra) 支持快速视频放大至 FullHD (1080p) 或 4K，同时保持清晰度。 +- **插值**：Apollo 模型用于帧插值，支持慢动作设置和自定义帧率（如 60fps）。 +- **动态压缩**：智能识别和处理视频中的压缩伪影，提高画面纯净度。 + +## 开始使用 Topaz Video Enhance + +1. 将 ComfyUI 更新到最新版本。 +2. 在节点列表中搜索 `Topaz Video Enhance` 节点。 +3. 连接视频输入并配置设置。 +4. 运行工作流。 + +## 技巧 + +- 将放大与[帧插值](/zh-CN/tutorials/utility/frame-interpolation)结合使用，获得更流畅、更高分辨率的输出。 +- 对于 AI 生成的视频，考虑在放大之前修复伪影，而不是依赖创意放大来修复它们。 +- 在处理完整视频之前，先在示例片段上测试不同的模型。 +- 对于需要严格保持原始保真度的内容，使用保守放大。