Skip to content
Snippets Groups Projects

Update racePromptPart.js

Open Update racePromptPart.js
Sheenariel/fc-pregmod:Sheenariel-pregmod-master-patch-06172 into pregmod-master
3 unresolved threads
Open Sheenariel requested to merge Sheenariel/fc-pregmod:Sheenariel-pregmod-master-patch-06172 into pregmod-master
3 unresolved threads

Expanded positive and negative for AI-generation. Excluding only Asian helps creating Caucasians while they are African/dark skinned.

Merge request reports

Members who can merge are allowed to add commits.
Loading
Loading

Activity

Filter activity
  • Approvals
  • Assignees & reviewers
  • Comments (from bots)
  • Comments (from users)
  • Commits & branches
  • Edits
  • Labels
  • Lock status
  • Mentions
  • Merge request status
  • Tracking
src/art/genAI/racePromptPart.js
5 5 positive() {
6 if (this.slave.race === "white") {
7 return "caucasian";
8 } else if (this.slave.race === "black") {
9 return "african";
6 if (this.slave.nationality == "stateless"){
7 if (this.slave.race === "white") {
8 return "Caucasian";
9 } else if (this.slave.race === "black") {
10 return "African";
11 } else if (this.slave.race === "asian") {
12 return "Asian";
13 }
14 return this.slave.race;
15 } else {
16 return "(" + this.slave.nationality + " origin:1.3)";
  • svornost
    svornost @svornost ·
    Reporter

    Replacing race with nationality falls apart very quickly for diverse nationalities. Think South Africa, Canada, the US, etc. Black and white South Africans simply can't use the same prompt if you expect good results, and neither can Anglo-Canadians, people from the First Nations, and Indo-Canadians.

  • Sheenariel
    Sheenariel @Sheenariel ·
    Author

    If you look only at the positive prompt, I agree. The negative part should add information to create a South African as Caucasian if they are really flagged as white. Everything I use does distinguish between origins, sometimes more sometimes less or hitting some pretty well and some not at all. It depends a lot on the model used. I would love to see a toggle in the options to use Race or Nationality for generation. But I do not even am able to create the export to test my fork nor am I even close to be deep enough in the sourcecode to know where to look and change everything needed.

  • Sheenariel changed this line in version 6 of the diff

    changed this line in version 6 of the diff

  • Please register or sign in to reply
  • svornost
    svornost @svornost started a thread on the diff
    • src/art/genAI/racePromptPart.js
    10 17 }
    11 return this.slave.race;
    12 18 }
    13 19
    14 20 /**
    15 21 * @returns {string}
    16 22 */
    17 23 negative() {
    18 if (this.slave.race !== "asian") {
    19 return "asian";
    24 if (this.slave.race === "white") {
    25 return "(African, Asian:1.5)";
    26 } else if (this.slave.race === "black") {
    27 return "(Asian, Caucasian:1.5)";
    28 } else if (this.slave.race === "asian") {
    29 return "(African, Caucasian:1.5)";
    • svornost
      svornost @svornost ·
      Reporter

      And this change has problems too (partly due to model overtraining). For example, if you don't have Asian as a negative, Pacific Islanders will all tend to look Japanese, which is just wrong.

    • Sheenariel
      Sheenariel @Sheenariel ·
      Author

      Well, I'd put "Asian" manually in the negative prompt if I think the model is putting to much emphasis on Japanese but this is looking at the parts of the prompts not the whole thing again.

      The positive should take care of it... should. No model is perfect but some are pretty good already.

      I am not masking my image generation by making a collage of several images so you can look at what I used.

      02645-74867571628375-RAW photo, (high detailed skin_1.2), 8k uhd, dslr, soft lighting, high quality, film grain, Fujifilm XT3, full body portrait, da.jpg02647-74867571628375-RAW photo, (high detailed skin_1.2), 8k uhd, dslr, soft lighting, high quality, film grain, Fujifilm XT3, full body portrait, da.jpg

    • svornost
      svornost @svornost ·
      Reporter

      If you use Batch Generation, the grid output actually contains all of the parameters in PNG info, just like each individual image does. You should be able to recover everything directly from the examples I posted, for example. image

      On the other hand, I don't know how to recover the parameters from the images you've posted...

      Here's a reference image of some actual pacific islanders:

      image

      And here we have woman, 18 years old, pacific islander, with the asian negative: image

      And the same positive prompt without the asian negative: image

    • Sheenariel
      Sheenariel @Sheenariel ·
      Author

      Hmmm... the one in your screenshot is not the same as you posted but the screenshot looks like A1111 which I am using as well. And I use the PNG Info a lot myself to go back to older results. With the 2 images from before, I only receive "Parameters None"

      Does gitgud maybe strip the information from posted images? When I try to use those I posted, not the ones on my HDD, I see no information either.

      As for the Pacific Islanders, I got a fitting result, I think, with the last tests before starting some more testing.

      Somehow my last comment down below is pending and I do not know what I might have done wrong ^^

    • svornost
      svornost @svornost ·
      Reporter

      Ugh, yeah, Gitlab strips image metadata. Fuck, that makes things harder.

      Let's see what happens if we attach instead of inlining:grid-0006

      Edit: And it just converts it to inline anyway. Sigh.

      Edited by svornost
    • Sheenariel
      Sheenariel @Sheenariel ·
      Author

      Nono, that one is working, just tested it

      So once again the pictures from the posts at the bottom and I did some testing as well and agree, at least with the model I am using right now, the weighting in the positive is not needed.

      Naturally it is all very spotty... changing the seed alone would change so many things.

      02681-78599260786103-RAW_photo___high_detailed_skin_1.2___8k_uhd__dslr__soft_lighting__high_quality__film_grain__Fujifilm_XT3__full_body_portrait__da02683-78599260786103-RAW_photo___high_detailed_skin_1.2___8k_uhd__dslr__soft_lighting__high_quality__film_grain__Fujifilm_XT3__full_body_portrait__da

      xyz_grid-0087-69360620553612-RAW_photo___high_detailed_skin_1.2___8k_uhd__dslr__soft_lighting__high_quality__film_grain__Fujifilm_XT3__full_body_portrait__da

      xyz_grid-0089-69360620553612-RAW_photo___high_detailed_skin_1.2___8k_uhd__dslr__soft_lighting__high_quality__film_grain__Fujifilm_XT3__full_body_portrait__da

      xyz_grid-0090-69360620553612-RAW_photo___high_detailed_skin_1.2___8k_uhd__dslr__soft_lighting__high_quality__film_grain__Fujifilm_XT3__full_body_portrait__daBut I would really like options to pick and change some more stuff. Like using the original version, my idea with the nationality and the combined version. As well as options for added weight. I am a programmer myself so I am aware this is not just simply done.

    • Please register or sign in to reply
    • svornost
      svornost @svornost ·
      Reporter

      I'd also warn against putting too much weight on race and nationality as parameters, because they are not independent of other parameters that you want to be able to adjust. Keep in mind, for instance, that the model has probably never seen an Egyptian woman with blue hair, but you want it to be able to generate them. Compare, for example:

      woman, 18 years old, egyptian origin, blue hair: image

      woman, 18 years old, (egyptian origin:1.5), blue hair: image

    • Sheenariel
      Sheenariel @Sheenariel ·
      Author

      I agree, 1.5 would be to much in the first place, though I only used 1.3 ... which did introduce still some elements.

      02641-74867571628375-RAW_photo___high_detailed_skin_1.2___8k_uhd__dslr__soft_lighting__high_quality__film_grain__Fujifilm_XT3__full_body_portrait__da02643-74867571628375-RAW_photo___high_detailed_skin_1.2___8k_uhd__dslr__soft_lighting__high_quality__film_grain__Fujifilm_XT3__full_body_portrait__da02631-74867571628375-RAW_photo___high_detailed_skin_1.2___8k_uhd__dslr__soft_lighting__high_quality__film_grain__Fujifilm_XT3__full_body_portrait__da02633-74867571628375-RAW_photo___high_detailed_skin_1.2___8k_uhd__dslr__soft_lighting__high_quality__film_grain__Fujifilm_XT3__full_body_portrait__da

      Edited by Sheenariel
    • svornost
      svornost @svornost ·
      Reporter

      Yeah, I used a slightly heavier weight so you could see what I meant more clearly. In your images, you can still see not just the extra elements being added, but also that black hair has started to dominate...that's an effect of the weight starting to mask another desired parameter. Weight has to be used really carefully, especially when your parameters aren't independent. When you're testing a specific parameter it can be really tempting to throw extra weight at it (you can imagine how someone wanting hair color to be followed more closely might throw some weight at it, after your changes...it just produces a constant arms race).

    • Sheenariel
      Sheenariel @Sheenariel ·
      Author

      Okay, found the issue with the "pending comment" ... hitting the big blue button for "review" out of reflex was the issue...

      It needs tinkering, not saying or claiming I am throwing the only and the perfect solution out there. I am used to giving the origin more weight so the models are not ignoring it. In the end, after I figured out how to get the whole thing on my computer and compile the html on my own for better testing, I could test it pretty well for the moment. I present to you, 2 times the exact same prompts, only the skin colors black vs. white are switched in the positive and negative.

      02681-78599260786103-RAW_photo___high_detailed_skin_1.2___8k_uhd__dslr__soft_lighting__high_quality__film_grain__Fujifilm_XT3__full_body_portrait__da02683-78599260786103-RAW_photo___high_detailed_skin_1.2___8k_uhd__dslr__soft_lighting__high_quality__film_grain__Fujifilm_XT3__full_body_portrait__da

      They have no added weight to them but come through more than the origin.

    • Please register or sign in to reply
  • Sheenariel added 1 commit

    added 1 commit

    • f63bc63e - Thrown in some additional help for models to not lose the Race variable.

    Compare with previous version

  • svornost mentioned in merge request !11579 (merged)

    mentioned in merge request !11579 (merged)

    • Please register or sign in to reply