CivArchive
    ← All articles
    Published September 16, 2024by AbstractPhila

    T5 non-technical block analysis.

    471 views8 reactions2 comments on CivitAI9 collected
    flux blocksML Researcht5block weight

    So I used this as a baseline. I left t5xxl_0 on and then checked each associative objects. t5xxl_0 seems to be a highly critical and important block, and I didn't quite get a good grip on what it does. Feel free to correct anything here and I'll test to fill in the new information.

    I used a node in ComfyUI named Flux Attention Seeker. This allowed me to slider up and slider down blocks in rapid succession. I ran a bunch of tests and gave each outcome a simplistic analysis. I ran 10 images and used my brain basically. They're likely wrong, but they definitely correlate heavily.

    I don't know for certain if any of this is accurate mind you.

    I haven't seen enough causal outcome to be certain. It seems that there is a mindful pattern based on training and finetuning and it can definitely be identified further.

    Through the entire test I left block_0 on while testing each of the others.

    I noticed the most anatomic destruction rather than creation, from blocks

    • block 20; associative anatomical structure and deformity, hand covering - censor 1, toggling off does not disable this, it does weaken it though

    • block 21; associative leg controller deformity causer, leg closer - censor 2, toggling this off causes different forms of deformity, so i assume this sort of thing isn't so simple

    Update 1:

    • block 21 appears to be at least somewhat responsible for text

    • Looks like every block has some loose or deep association with text. Some damage more than others, but mostly simple text stays intact until you turn off seemingly enough of them. More akin to lobotomy than anything.

    • Simple text can be generated using only block 0 to block 7.

    If I could get some contrasting checks and balances that would help.

    t5xxl_1 analysis;

    block 0; face, subject, single object concept

    block 1; associative base with other objects

    block 2; seems to have added leg context

    block 3; additional details and context

    block 4; added additional and counter-intuitive object interactions

    block 5; enabled hand association? vibes heavily with hagrid

    block 6; effects and liquid

    block 7; full locomotive pose and body

    block 8; hand motion and hand poses

    block 9; situational and background context

    block 10; age?

    block 11; scale and size

    block 12; muscle and skin tone definition finetunining

    block 13; material and reflective effects

    block 14; object to object assocation

    block 15; surface interaction

    block 16; object blending - clothes

    block 17; multi dimensional geometry, room corners, stairs, etc

    block 18; gravity

    block 19; rotational axis

    block 20; associative anatomy and deformity causer - censor 1

    block 21; associative leg controller deformity causer - censor 2 - text?

    block 22; multi humanoid subject association

    block 23; associative censoring, group censor?