[For Advanced Anima Users] I'm releasing the workflow I use as my main one for free!
Good evening, I'm HKMC, the guy who loves Anima.
This time, I'm releasing the workflow I use.
*8/15
Updated the version.
It looks complicated at first glance, but what it's doing isn't that difficult.
I'm just taking a few detours.
I'm not good at using subgraphs, so the workflow is a bit messy. Sorry about that.
Workflow flow and basic specifications
Base generation
↓
1st HiresFix
↓
2nd HiresFix
↓ Branch
Sampling by material
↓
Each Detailer
It's a simple flow like that.
You can toggle the bypass for each process group using the nodes below.
If you set everything to 'no', it will only perform base generation.
Base generation
The base generation part is cluttered, but I'm not doing anything special.
I'm just loading a LORA and generating.
It includes a custom node I made called 'Anima Resolution Selector'.
It's just a preset version of a standard latent image node, so please change it to your own.
The Step value is connected to a Math Expression and a floating-point number, which automatically calculates the number of steps.
By default, it is set to the number of steps configured in the base sampler multiplied by 0.3.
This ratio is the golden ratio that doesn't break the original image.
1st HiresFix
Upscale without an upscale model using a node called DaSiWaTorchResize
You can ignore the target_width and target_height values. They are bypassed.
The scale_multiplier value determines the scale factor.
↓
Detail adjustment with Detail Daemon Sampler
The detail_amount value is the strength.
start and end specify at which point in the steps the daemon should take effect and end.
It's better not to touch the others.
↓
Adjust the noise removal curve of the scheduler
You can adjust how the scheduler's noise removal decreases using Sigmas Power.
If you don't know much about noise removal, it's fine to leave it at the default of 1.0.
You can check the curve as an image with SigmasPreview.
↓
Resampling
↓
To 2nd HiresFix
2nd HiresFix
Upscale without an upscale model using a node called DaSiWaTorchResize
↓
Automatically mask only the character.
(Since I added strong details to the whole image in the first pass, I'm only adjusting the character here)
BBOX Detector (SEGS) and SAMDetector (combined)
I'm slightly expanding the mask range with the dilation value.
This is to avoid breaking the lines by pushing the edges too hard.
*8/11
Added a toggle switch for the second HiresFix to apply to the whole image, character only, or background only.
↓
Adjust the noise curve of the scheduler
↓
Resampling only the masked area (character)
↓
Adjust color tone by referencing the image from the base generation using Image Color Match.
↓
To Face&Eyes Detailer (or to sampling by material)
Branching: Material-specific sampling
Input the material you want to mask into Florence2Run to automatically mask it.
↓
In the masked image composition node, use the image data from before sampling as a reference to composite layers for areas that were over-adjusted during the second HiresFix.
Destination: Original image
Source: Reference image
↓
Adjust and sample the composited areas.
↓
Use Image Color Match to adjust the color tone by referencing the image from the base generation.
↓
To Face & Eyes Detailer
Face & Eyes & Hand Detailer
Automatically identify and re-sample faces, eyes, and hands.
↓
End
That is the flow.
Works created using this workflow process
Improvements from last time
Last time, there was an issue where the texture of the character's leather was lost during the second HiresFix.
To fix that, I have newly incorporated a branching material-specific sampling process this time.
By simply entering "corset" into the Florence2Run text input field, the AI automatically masks the corset.
I created a process where the corset image from when the texture was good is composited as a reference, then sampled to blend it in.
Below is the result.
Huge success!!
Finally
It's a pretty cluttered workflow, but I think it's quite easy to use (since I'm the one who made it lol).
If there are any strange parts, please let me know.
I will not accept suggestions to just use inpainting (ヽ´ω`)
Me, inpaint, hate...
Please give it a try!
※Updated 8/11
The workflow above is fine, but I have corrected the second HiresFix section.
I have added a toggle switch for the second HiresFix to choose whether to:
apply it to the whole image,
apply it only to the character using a mask,
or apply it only to the background.
Setting the select value of the Switch(Any) node to 1 runs HiresFix on the whole image, 2 runs it only on the character, and 3 runs it only on the background.
