Upgrading the Anima workflow and distributing custom nodes.
Good evening, this is HKMC, feeling a bit down because the new 2.9B parameter base model for Anima just got heavier without any real changes.
Today, I am announcing and distributing an upgraded version of the Anima workflow I introduced previously.
Along with this, I will also be distributing the custom node used in the workflow,
“Anima Resolution Selector”
, and explaining how to use it briefly.
The previous article introducing the workflow can be found here.
Changes from V1.0 to V1.1
The changes from V1.0 to V1.1 are as follows.
Changed the sampling method for the second HiresFix to a selectable option.
Changed the sampling method for material-specific sampling to a selectable option.
Separated the positive prompt for material-specific sampling.
Added an Eyes Detailer step. (Technically, it was separated from Face Detailer.)
Removed Detail Deamon from the base generation step.
Layout & numerical adjustments.
These are the changes.
① For the second HiresFix, I made it possible to select
1. Apply to the entire illustration, 2. Apply only to the character, or 3. Apply only to the background
using a switch. Please see the notes within the workflow for details. It's just a click away.
② Previously, material-specific sampling was a process where the image after the second HiresFix was used as the base image, and the image before sampling was used as the source image to composite and adjust. In addition to that, I added a process to sample the selected mask directly without compositing, and made it selectable via a switch. Please see the notes within the workflow for details. It's just a click away.
■ Supplementary explanation of material-specific sampling ■
I assume you might not understand what I'm saying, so I'll explain it clearly. Material-specific sampling is a process to correct the image created by the second HiresFix. For example, suppose there is a part of the image after the second HiresFix that you don't like. For instance, you don't like the shape or color of a pendant. When you input 'Pendant' into the Florence2Run node, the AI analyzes the image and masks the pendant. This process involves either compositing the masked area with the image before the second HiresFix sampling (after upscaling) to blend it, or sampling the masked area anew. In other words, it's a process similar to inpainting.
③ Since I added a process to sample directly without mask compositing in the material-specific sampling step, I added a new independent positive encoder (prompt).
④ I added the EyesDetailer, which was highly requested (0 requests). This increases the detail of the eyes and makes the eyeballs sparkle.
⑤ I removed Detail Deamon from the base generation. Please note that it remains in the first HiresFix. If you do not want to use Detail Deamon, you can disable it by setting the detail_amount value to 0, so if bypassing is a hassle, just set it to 0.
⑥ I adjusted the layout and the numerical values for the sampling items. These are my personal preferences, so please feel free to change them as you like.
Distribution of the custom node [Anima Resolution Selector]
It has a cool name, but it's nothing special.
It is simply a node that turns latent images (image sizes) into presets.
You can edit the image size presets yourself.
While there are similar nodes among existing custom nodes, there were ones branded with SDXL but none branded with [Anima], so I made it for my own satisfaction.
I've been using it for over two months now, and I haven't encountered any bugs at all!
How to use Anima Resolution Selector.
Simply place the unzipped ComfyUI-Anima-Resolution folder into the ComfyUI Custom_nodes folder.
Inside the ComfyUI-Anima-Resolution folder, there is a JSON file called presets so please open it with Notepad or similar and set your preferred image sizes as presets.
The following image sizes are described by default.
"Portrait 2:3 (832x1248)": [832, 1248],
"Portrait 3:4 (912x1216)": [912, 1216],
"Portrait 4:5 (1024x1280)": [1024, 1280],
"Portrait 9:16 (720x1280)": [720, 1280],
"Landscape 4:3 (1216x912)": [1216, 912],
"Landscape 3:2 (1248x832)": [1248, 832],
"Landscape 5:4 (1280x1024)": [1280, 1024],
"Landscape 16:9 (1280x720)": [1280, 720],
"Square (1040x1040)": [1040, 1040]
Portrait is vertical, Landscape is horizontal, and square is square.
Since it only references the values inside [ ], you can write whatever you like inside the " ".
You can select either preset or custom in the mode.
Please rest assured that when you select preset, the custom_width and custom_height values are ignored.
When you select custom, it references the custom_width and custom_height values.
hires_scale is the upscale factor item. (1.3x by default)
Please rest assured that the values are not applied just by touching them.
They are only applied when you link the output hires_width and hires_height.
By connecting info to Show Text, you can check the image size and the size after scaling in advance.
Since you can turn the image sizes you use often into presets and it has high versatility, it's surprisingly convenient and a cute little tool, so please treat it well.
If you encounter any bugs, please report them in the comments 🙏
Ciao!