Anima Tiled Upscale Workflow

This ComfyUI workflow is designed for Anima tiled upscaling, high-resolution image enhancement, and tile-based detail reconstruction. The main purpose of this workflow is to take an existing image, enlarge it, split it into manageable tiles, automatically describe each tile with Florence2, refine each tile through Anima Preview, and then reconstruct the final high-resolution image with improved clarity and detail.

Unlike a simple one-click upscale workflow, this graph uses a multi-stage enhancement structure. It does not only increase image size. It first uses a traditional upscale model to enlarge the source image, then uses AI-based tile refinement to rebuild details locally. This makes the workflow useful for images that need more texture, cleaner edges, stronger local detail, and better final publishing quality.

The workflow is built around Anima Preview, using anima-preview.safetensors as the main diffusion model. It also uses qwen_3_06b_base.safetensors as the text encoder and qwen_image_vae.safetensors as the VAE. The workflow includes an optional LoRA route through LoraLoader, which can be used to apply a specific style, model feature, or subject-related enhancement during the tile refinement stage. In the uploaded graph, the LoRA example is loaded with model and clip strengths around 0.85.

The first stage starts with LoadImage. The user uploads the image that needs to be enlarged and refined. This can be an AI-generated image, anime artwork, character illustration, concept art, product-style render, social media cover, or any image that already has a good composition but needs higher resolution and better detail.

The image is then passed into ImageUpscaleWithModel using 4x_NMKD-Siax_200k.pth. This is the first-stage traditional upscaler. It enlarges the image while preserving the original layout and structure. Traditional upscalers are stable and fast, but they often cannot add enough semantic detail. That is why this workflow continues with Anima-based tile refinement after the initial enlargement.

After the first upscale, ImageScaleToTotalPixels is used to control the final working size. In the uploaded setup, the workflow uses Lanczos scaling and targets around 3 megapixels. This gives users a practical way to control output scale without manually calculating width and height. It also helps keep the workflow manageable before entering the tiled refinement stage.

The next key stage is tile splitting. The workflow uses TTP_Tile_image_size to calculate tile dimensions based on the image size, width factor, height factor, and overlap rate. In the uploaded graph, the width factor and height factor are set to 2, with an overlap rate around 0.35. TTP_Image_Tile_Batch then splits the image into tile batches, with a tile size route around 1024 x 1024. This allows the workflow to process large images more safely and with better local detail.

Tile-based processing is important because high-resolution images can be too heavy for a full-frame diffusion pass. If the entire image is processed at once, VRAM usage may become too high, and the model may not focus enough on local details. By splitting the image into tiles, the workflow gives Anima Preview smaller and more focused regions to refine.

The workflow then converts the tile batch into an image list and sends each tile into Florence2. Florence2Run is configured with a detailed caption task using the Florence-2-Flux-Large model. This stage automatically generates a description for each tile. The captions are displayed through ShowText, allowing users to see what Florence2 detected in each region.

This automatic captioning stage is one of the most useful parts of the workflow. Different tiles may contain different content. One tile may contain a face, another may contain clothing, another may contain a background object, and another may contain a vehicle or landscape. Instead of using one generic prompt for the whole image, Florence2 gives the workflow local semantic descriptions for each tile. These descriptions are then used as positive conditioning for Anima refinement.

The positive conditioning is created through CLIPTextEncode using the Florence2-generated captions. The negative prompt is also encoded through CLIPTextEncode. The included negative prompt suppresses common image problems such as low quality, worst quality, blur, bad anatomy, bad hands, extra fingers, fused fingers, deformed faces, text, watermark, logo, and JPEG artifacts. This is especially useful for tile refinement because local redraw can sometimes introduce unwanted artifacts if not constrained.

The tile refinement stage uses VAEEncode, KSampler, and VAEDecodeTiled. Each tile is encoded into latent space through the Qwen image VAE, then processed with Anima Preview. In the uploaded setup, the KSampler uses around 31 steps, CFG 4, er_sde sampler, simple scheduler, and a denoise value around 0.25. This is a conservative low-denoise setting, suitable for detail enhancement rather than full regeneration.

The low denoise value is important. In an upscale workflow, the goal is usually to preserve the original image while improving quality. If denoise is too high, the model may change faces, clothing, composition, background objects, or character identity. If denoise is too low, the result may not gain enough detail. A value around 0.25 is a good starting point for faithful enhancement.

After sampling, VAEDecodeTiled decodes the refined latent tiles. The tiled decode settings use a large tile size and overlap, which helps manage memory and reduce decoding artifacts. The decoded image list is converted back into a batch through easy imageListToImageBatch. Then TTP_Image_Assy reconstructs the full image from the processed tiles, using position data, original size, grid size, and padding. In the uploaded setup, padding is set around 128 to help reduce visible seams between tiles.

The final output is saved through SaveImage. The workflow also includes Image Comparer, allowing users to compare the original image and the final reconstructed image in a slide view. This is very useful for judging whether the enhancement is actually successful. A good upscale result should improve sharpness, texture, line clarity, and local detail without changing the original image identity too much.

This workflow is especially useful for anime-style images and stylized artwork. Anima Preview is well suited for anime and illustration refinement, while Florence2 helps generate local descriptions for each tile. This combination allows the workflow to enhance faces, hair, clothing folds, mechanical details, background objects, and environmental textures more intelligently than a simple pixel upscaler.

It can also be used for Civitai showcase images, RunningHub workflow examples, YouTube thumbnails, Bilibili covers, AI character art, concept art, poster-style images, and social media visuals. When an image already looks good but lacks final-resolution polish, this workflow can act as a finishing pass.

Main features:

- Anima tiled upscale workflow

- Uses anima-preview.safetensors

- Qwen 3 0.6B text encoder support

- Qwen image VAE support

- Optional LoRA route for style or subject enhancement

- First-stage 4x_NMKD-Siax upscaling

- ImageScaleToTotalPixels output size control

- TTP tile size calculation

- TTP_Image_Tile_Batch tile splitting

- Florence2 automatic tile captioning

- Tile-specific positive prompt generation

- Low-denoise Anima refinement

- VAEDecodeTiled for memory-friendly decoding

- TTP_Image_Assy final reconstruction

- Image Comparer for before-and-after inspection

Recommended use cases:

Anime image upscaling, Anima Preview testing, tiled image refinement, character art enhancement, AI illustration polishing, concept art cleanup, high-resolution cover preparation, Civitai showcase image output, RunningHub workflow publishing, Bilibili cover image preparation, YouTube thumbnail enhancement, poster output, detail reconstruction, and before-after comparison demonstrations.

Suggested workflow:

Start by loading a clean source image. The workflow works best when the original image already has a strong composition and clear subject. It can improve softness and detail, but it cannot fully fix a source image that is extremely broken, heavily compressed, or structurally wrong.

Run the first-stage upscale with 4x_NMKD-Siax. This gives the workflow a larger base image before Anima refinement. Then use ImageScaleToTotalPixels to control the target resolution. If the output is too large for your GPU, reduce the megapixel value before entering the tile stage.

Use the tile settings carefully. The uploaded setup uses a tile workflow around 1024 x 1024 with overlap support. If seams appear in the final output, increase overlap or padding. If VRAM usage is too high, reduce the final megapixel target or use smaller tile settings.

Let Florence2 generate captions for the tiles. Check the ShowText output when possible. If Florence2 describes a tile incorrectly, you can manually adjust the prompt or simplify the image before processing. Better tile captions usually produce more accurate local refinement.

Keep denoise conservative for faithful upscaling. The default low-denoise route is designed to preserve the original image while improving local detail. Increase denoise only if you want stronger redraw. Reduce denoise if faces, hands, clothing, or identity begin to change too much.

Use the negative prompt to suppress common artifacts. For anime and illustration, keep terms like low quality, blurry, bad anatomy, bad hands, extra fingers, deformed face, text, watermark, and logo. For realistic images, you may add more realism-specific artifact controls.

After reconstruction, use Image Comparer to inspect the before-and-after result. Look at face stability, line clarity, hair detail, fabric texture, object edges, background continuity, and whether any tile seams are visible. A good result should feel sharper and richer while still looking like the same image.

For character images, inspect the face and hands first. For mechanical or product-style images, inspect edges, logos, surface texture, and symmetry. For background-heavy images, inspect tile borders and repeated patterns carefully.

This workflow is designed as a practical Anima high-resolution finishing pipeline for ComfyUI users. It combines traditional upscaling, automatic tile captioning, Anima low-denoise refinement, tiled decoding, final reconstruction, and before-after comparison into one usable graph. It is especially useful for creators who need clean, detailed, and publishable images for Civitai, RunningHub, YouTube, Bilibili, and social media platforms.

🎥 YouTube Video Tutorial

Want to know what this workflow actually does and how to start fast?

This video explains what the tool is, how to launch the workflow instantly, and shares my core design logic — no local setup, no complicated environment.

Everything starts directly on RunningHub, so you can experience it in action first.

👉 YouTube Tutorial: https://youtu.be/J2A8JWDCUhk

Before you begin, I recommend watching the video thoroughly — getting the full context helps you understand the tool faster and avoid common detours.

⚙️ RunningHub Workflow

Try the workflow online right now — no installation required.

👉 Workflow: https://www.runninghub.ai/post/2021929692445544449/?inviteCode=rh-v1111

If the results meet your expectations, you can later deploy it locally for customization.

🎁 Fan Benefits: Register to get 1000 points + daily login 100 points — enjoy 4090 performance and 48 GB super power!

📺 Bilibili Updates (Mainland China & Asia-Pacific)

If you’re in the Asia-Pacific region, you can watch the video below to see the workflow demonstration and creative breakdown.

📺 Bilibili Video: https://www.bilibili.com/video/BV1FscqzREni/

☕ Support Me on Ko-fi

If you find my content helpful and want to support future creations, you can buy me a coffee ☕.

Every bit of support helps me keep creating — just like a spark that can ignite a blazing flame.

👉 Ko-fi: https://ko-fi.com/aiksk

💼 Business Contact

For collaboration or inquiries, please contact aiksk95 on WeChat.

🎥 YouTube 视频教程

想了解这个工作流到底是怎样的工具，以及如何快速启动？

视频主要介绍 工具定位、快速启动方法 和 我的构筑思路。

我们会直接在 RunningHub 上进行演示，让你第一时间看到实际效果。

👉 YouTube 教程： https://youtu.be/J2A8JWDCUhk

开始前建议尽量完整地观看视频 —— 把握整体思路会更快上手，也能少走常见弯路。

⚙️ 在线体验工作流

现在就可以在线体验，无需安装。

👉 工作流： https://www.runninghub.ai/post/2021929692445544449/?inviteCode=rh-v1111

打开上方链接即可直接运行该工作流，实时查看生成效果。

如果觉得效果理想，你也可以在本地进行自定义部署。

🎁 粉丝福利： 注册即送 1000 积分，每日登录 100 积分，畅玩 4090 体验 48 G 超级性能！

📺 Bilibili 更新（中国大陆及南亚太地区）

如果你在中国大陆或南亚太地区，可以通过下方视频查看该工作流的实测效果与构思讲解。

📺 B站视频： https://www.bilibili.com/video/BV1FscqzREni/

我会在 夸克网盘 持续更新模型资源：

👉 https://pan.quark.cn/s/20c6f6f8d87b

这些资源主要面向本地用户，方便进行创作与学习。