Strict mirroring版本IL V1 (ID: 1266826)

This is early experimental LoRa for researching borders of training the concept of strict mirror.

I think it is possible because SDXL can draw mirror floors.

Status

IL V2: this LoRa version can increase probability (+24% for 8-20 steps) of getting strict reflections in a new generation. Increasing steps is refining the results. Increasing LoRa's weight can refine the results (60% success with weight=2.0).

IL V1: this LoRa version can slightly increase probability (+10% for 4 dmd2-steps, +12-16% for 8 dmd2-steps) to find strict reflections in a new generation. Increasing steps is refining the results.

I go to the next step of the experiment.

Changelog

IL V2

• Added negative tokens:

  • x_out_of_mirror (protects against a reflection outside mirror's bounds)

  • x_differect_reflection (supress a wrong symmetry)

  • different reflection (same the x_out_of_mirror but danbooru tag)

• Added ai-generated images with negative tokens to dataset.

• 60 epochs instead of 10 epochs of version 1

• Trained with cleaned dictionary: x_mirror_reflection, x_reflection_face, x_reflection_facing_aside, x_reflection_facing_away, reflection, mirror, hand on mirror, full-length mirror, broken mirror, reflection focus, looking at mirror.

Usage

Connect LoRa in your workflow and then activate by word x_mirror_reflection in your prompt. Next choose one of these controls for the position of the character's face relative to the mirror:

• x_reflection_facing_away

• x_reflection_facing_aside

• x_reflection_face

Now you can use negative prompt (+8% to find good mirroring in IL V2): x_out_of_mirror, x_differect_reflection, different reflection

描述:

1st iteration.

I used DMD2 and SPO LoRas as accelerators, TCD sampler, sgm uniform as scheduler, [holy mix] Illustrous model.

It isn't strict enough but I hate to wait.

Metrics:

-------------------------------------

4 steps, with lora, 100 runs

prompt:

-------------------------------------

artist, location, character,

masterpiece, best quality, amazing quality, very aesthetic, absurdres, newest,

(safe rating),

x_mirror_reflection, {x_reflection_face, {|touching mirror}|x_reflection_facing_aside|x_reflection_facing_away},

mirror, reflection, {1girl|1girl|__character__, 2girls}, symmetrical reflection, pose is symmetrical for the reflection,

{dynamic pose|lazy pose|sitting}

-------------------------------------

total:

14% - strict reflection

9% - strict reflection but outside mirror borders

51% - wrong symmetry

26% - trash (prompt misunderstanding)

-------------------------------------

-------------------------------------

4 steps, with lora, 100 runs

same prompt

-------------------------------------

total:

4% - strict reflection

6% - strict reflection but outside mirror borders

21% - wrong symmetry

69% - trash (prompt misunderstanding)

-------------------------------------

Hypothesis: low steps is not enough for generate sophisticated concepts.

-------------------------------------

8 steps, with lora, 25 runs

same prompt

-------------------------------------

total:

28% - strict reflection

16% - strict reflection but outside mirror borders

32% - wrong symmetry

24% - trash (prompt misunderstanding)

-------------------------------------

-------------------------------------

8 steps, without lora, 25 runs

same prompt

-------------------------------------

total:

16% - strict reflection

4% - strict reflection but outside mirror borders

24% - wrong symmetry

56% - trash (prompt misunderstanding)

-------------------------------------

Hypothesis: prompt generates too many trash due to "symmetrical" tokens.

-------------------------------------

8 steps, with lora, 25 runs

prompt:

-------------------------------------

masterpiece, best quality, amazing quality, very aesthetic, absurdres, newest,

(safe rating),

x_mirror_reflection, {x_reflection_face, {|touching mirror}|x_reflection_facing_aside|x_reflection_facing_away},

mirror, reflection, {1girl|1girl|__character__, 2girls}

{dynamic pose|lazy pose|sitting}

-------------------------------------

total:

28% - strict reflection

12% - strict reflection but outside mirror borders

56% - wrong symmetry

4% - trash (prompt misunderstanding)

-------------------------------------

-------------------------------------

8 steps, without lora, 25 runs

same prompt

-------------------------------------

total:

12% - strict reflection

0% - strict reflection but outside mirror borders

16% - wrong symmetry

72% - trash (prompt misunderstanding)

-------------------------------------

Conclusions: this LoRa version slightly increase probability (+10% for 4 steps, +12-16%) to find strict reflections in a new generation. Increasing steps is refining the results.

I go to the next step of the experiment.

训练词语: x_mirror_reflection,x_reflection_facing_away,x_reflection_facing_aside,x_reflection_face