cowpuncher/Krita_AI_Addon_Tools / docs /Anima Reference.txt
cowpuncher's picture
download
raw
8.79 kB
ANIMA REFERENCE 1.1 (for AI Image Diffusion 1.53.0)
=====================================================
Makes three of Krita's own control layers work with Anima styles: the same
"Add Control Layer" rows you use with Illustrious. Without it Krita shows
"Not supported for Anima" there.
Reference a character. Two Reference layers = left and right character.
Style like Reference, and the art style of the picture is taken too.
Face one more picture of the same character (a face close-up, the outfit
from another side ...). Up to 5 pictures in total.
Behind the scenes it uses two ComfyUI nodes: "Anima Ref Adapter" by ElectronGirl
and "SnapRef". No file of the AI plugin is changed; the add-on hooks in when Krita
starts and checks first that the plugin is the version it knows.
Check that everything is in place: Tools > Scripts > Anima Reference
(wait until the AI docker shows it is connected). It lists what was found on the
server and what the last job used.
ONE CHARACTER
-------------
1. Put the reference picture on its own layer. Any size. Transparent areas are
treated as white. Only the part of the layer inside the canvas is read.
(The AI Reference Board does this step for you.)
2. Pick an Anima style in the AI docker.
3. Click "Add Control Layer". In the new row choose "Reference" in the first
dropdown and your reference layer in the second.
4. Write a prompt that describes the SCENE, not the character:
masterpiece, best quality, score_7, safe, 1girl, solo,
cowboy shot, sitting, cafe, holding cup, smile, indoors, day
Leave out the character's name, artist tags and most hair / eye / outfit
tags - the reference supplies those. Ask for a pose that differs from the
reference picture.
5. Generate.
The reference layer can be hidden (eye icon); it is still used.
A chest-up picture carries no legs or shoes: use a full-body picture if the outfit
matters.
THE ART STYLE TOO
-----------------
Set the row to "Style" instead of "Reference". The character and the look of the
picture (line, colouring, shading) both come from it. With "Reference" the look
comes from your checkpoint and prompt.
The model reads all pictures of a job together, so one Style row switches the art
style on for every picture in that job. There is no "style only, no character".
MORE PICTURES OF THE SAME CHARACTER
-----------------------------------
Keep one Reference (or Style) row for the main picture and add the others as
"Face" rows: for example a full-body picture as Reference and a face close-up as
Face. Up to 5 pictures. Keep "solo" in the prompt.
Each row has its own slider: lower a Face row if that picture changes the look
too much (position 1 or 2), raise it if it is ignored.
The more pictures, the smaller each one is read: 1 or 2 pictures at up to 1
megapixel each, 3 at 0.67, 4 at 0.5, 5 at 0.4.
TWO CHARACTERS
--------------
Add two Reference rows (or Style rows). The first row is the LEFT character, the
second the RIGHT one. In the prompt say who is where and use 2girls / 1boy, 1girl
instead of solo:
2girls, the girl on the left is a policewoman with black hair, the girl on
the right is a mecha girl with aqua hair, standing side by side, city, day
A wide picture works best. One picture per character: Face rows cannot be combined
with two characters, and more than two characters are refused with a message.
THE STRENGTH SLIDER
-------------------
The small slider in the control layer row has five positions. What they do depends
on which node handles the job (shown under "Last job" in Tools > Scripts > Anima
Reference).
Reference rows, Anima Ref Adapter:
position strength ref_guidance push stops at
1 (lowest) 0.4 2.4 60%
2 0.7 3.45 80%
3 (default) 1.0 4.5 100% <- its author's setting
4 1.2 5.2 100%
5 (highest) 1.4 5.9 100%
Two rows: the two strengths are averaged. Turbo styles (CFG 1): positions 1 to 3
all give ref_guidance 1, position 4 gives 1.7, position 5 gives 2.4.
Style and Face rows (and Reference rows when SnapRef handles them):
position weight of that picture
1 (lowest) 0.4
2 0.7
3 (default) 1.0 <- as trained
4 1.2
5 (highest) 1.4
Every picture has its own weight. Two characters share one (the average).
"Range" is not used. Strength 0 (under "Use custom values") leaves the picture out.
Raise the slider when the character drifts away from the reference. Lower it when
colours get too strong or the pose of the reference is copied.
WHICH NODE IS USED
------------------
Reference rows only Anima Ref Adapter
any Style or Face row SnapRef
Regions in the document Anima Ref Adapter (Reference rows only, see below)
Option "Use SnapRef for Reference layers too" (Tools > Scripts > Anima Reference):
plain Reference rows then go through SnapRef as well. By its authors' numbers SnapRef
costs about 15% more time than no reference, the Anima Ref Adapter about twice the
time at the default slider position (832x1216: 33 s -> 65 s on an RTX 3060 Ti).
The two do not draw the same picture; try both and keep what suits your checkpoint.
REGIONS
-------
Reference rows work together with Krita Regions (also one Reference row per
region: the regions are read left to right). The reference then runs at its basic
strength and the slider has no effect.
Style and Face rows cannot be used while the document has two or more regions;
the add-on says so instead of starting the job. SnapRef and Krita's region node
both replace the same part of the model, and running them in one job stops with
an error on the server.
GOOD TO KNOW
------------
* Selections, Refine (strength below 100%) and the Upscale workspace work the same
way. Filling a hole with an empty prompt adds no reference on its own.
* Memory: SnapRef keeps about 0.9 GB of graphics memory per megapixel of reference
(its authors' figure), so five pictures cost about as much as two.
* Both adapters were trained on Anima base v1.0. Fine-tunes and merges vary:
identity usually carries over better than small details (scars, earrings,
patterns).
* Style and Face are new in 1.1. They were run through every kind of job on a test
server, but with stand-in weights: how good the pictures are on your checkpoint
is for you to judge. If they look off, lower the slider or use plain Reference.
* Composition stays unsupported for Anima.
IF SOMETHING IS OFF
-------------------
"Anima Ref Adapter is missing on the server (run install.bat)" or
"SnapRef is missing on the server (run install.bat and answer Y to SnapRef)"
in the control layer row: the server does not have that node or its weights.
Close Krita, run install.bat again, start Krita.
Still "Not supported for Anima":
the add-on is not switched on. Settings > Configure Krita > Python Plugin
Manager > tick "Anima Reference (add-on for AI Image Diffusion)" > restart.
A message "Anima Reference add-on ... NOT ACTIVE" a few seconds after Krita starts:
the AI plugin is a different version than this add-on was made for. Nothing is
changed in that case; everything else in the AI plugin works as before.
"Anima can show two different characters at most ..."
three or more Reference/Style rows. Set the extra pictures of the same
character to Face.
"Face layers ... only work with a single Reference or Style layer."
two characters plus a Face row. Use one picture per character.
"Style and Face layers can't be combined with Regions on Anima ..."
see REGIONS above.
After "Upgrade" of the server from Krita's settings the rows say a node is missing:
a server upgrade removes extra nodes. Run install.bat again.
The server stops a job with 'NoneType' object is not subscriptable :
SnapRef and the region node met in one job (for example through a workflow
of your own). Restart Krita, which restarts its server.
Result looks like a normal Anima picture, reference ignored:
raise the slider; check the style is an Anima style and the row points at
the right layer.
Colours burnt / too strong: lower the slider one or two positions.
Out of memory: fewer pictures, or a smaller canvas or selection.
REMOVE
------
Python Plugin Manager: untick "Anima Reference", restart Krita. Krita is then
exactly as before. Files: see README.txt, "Remove".

Xet Storage Details

Size:
8.79 kB
·
Xet hash:
23fc9ce9310dd7b712d2c566498efa0564c2b74929c98985b1cbe48269c3a049

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.