Quick question regarding the @perchance image generator.

Today when attempting to generate images any details entered after approximately 300 words or 1,900 characters appears to be ignored or dropped by the generator. This was not the case a few days ago.

Is this expected behaviour? I’ve been told the image generator was updated in the last couple of days.

  • RudBo@lemmy.world
    link
    fedilink
    English
    arrow-up
    4
    arrow-down
    1
    ·
    2 months ago

    each about 50 Words

    • [0-50 Words: Red Bicycle]
    • [50-100 Words: Teddy Bear]
    • [100-150 Words: Soccer Ball]
    • [150-200 Words: Rainbow Wall]
    • [200-250 Words: Telephone Booth]
    • [250-300 Words: Woman]
    • [300-350 Words: Rottweiler Dog]
    • [350-400 Words: Yellow Excavator]
    A vintage style red bicycle is parked on the pavement. It features bright cherry red paint that is slightly chipped in places, chrome handlebars that are gleaming in the sun, a worn brown leather saddle, and a wicker basket on the front filled with some dried wildflowers. The tires are thin and black.
    
    Leaning directly against the front wheel spokes of that red bicycle is an old, well-loved teddy bear. Its brown fur matted and worn thin in spots from years of hugging, one black button eye slightly loose hanging by a thread, and it wears a faded blue wool scarf wrapped tightly around its neck.
    
    Just a few feet away from the bear, resting on the cracked asphalt surface, is a classic black and white hexagonal panel soccer ball. It looks used, scuffed with dirt stains and green grass marks from a recent game. The synthetic leather surface is worn smooth in the areas where it has been kicked repeatedly.
    
    Serving as the backdrop for these items is a massive brick wall covered entirely in vibrant, chaotic rainbow graffiti art. Thick dripping stripes of red, orange, yellow, green, blue, and purple spray paint run down the textured bricks, overlapping and swirling into complex patterns that cover every single inch of the rough masonry surface.
    
    Standing prominently in front of that painted wall is an iconic, British-style bright red cast-iron telephone booth. The glass panels are slightly dusty, and the heavy translucent TELEPHONE sign is illuminated from within. Inside the booth, can clearly see an old rotary phone apparatus and a tattered phone book hanging by a metal chain.
    
    Standing on the roadside curb near the booth is a woman in her thirties with a quiet, pensive expression. She is wearing a beige trench coat over blue denim jeans, looking down at her glowing smartphone screen. A gentle breeze blows her shoulder-length brown hair across her face as she stands still waiting for someone.
    
    Sitting patiently right next to the woman's leather boots is a large, muscular Rottweiler dog with a glossy black and tan coat. It is wearing a thick leather collar studded with sharp silver metal spikes and is attached to a heavy-duty steel chain leash that the woman is holding quite loosely in her hand.
    
    Finally, looming in the background behind the entire scene is massive heavy construction machinery, specifically a bright yellow hydraulic excavator digger. Its giant metal claw bucket is resting on a large pile of dirt, and the caterpillar tracks are covered in dried mud, suggesting active construction work is happening nearby on this busy street scene.
    
    • Almaumbria@lemmy.world
      link
      fedilink
      English
      arrow-up
      6
      ·
      2 months ago

      It is remarkably attentive, until the cutoff point at which it will simply omit elements. I could be wrong, but my gut tells me the latter arose as an unintended side effect of implementing the former.

      • A_A@lemmy.world
        link
        fedilink
        English
        arrow-up
        2
        ·
        2 months ago

        Someone reuse part of your prompt :
        ( + mods … 488 words)
        the result, like you suggested, is that the model doesn’t simply truncate at 300 words or tokens

  • yotv@lemmy.world
    link
    fedilink
    English
    arrow-up
    5
    arrow-down
    2
    ·
    2 months ago

    I hope this is just a glitch because this update so far has been a total downgrade no improvements at all. its better to just revert to the previous version. i was a big fan of the update they did in april 2025 and that one had some glitches too. but this doesn’t seem to do anything the previous version couldn’t and just has more limitations .

  • yotv@lemmy.world
    link
    fedilink
    English
    arrow-up
    2
    arrow-down
    2
    ·
    2 months ago

    for me its way worse right now it already starts skipping things at 20 words

  • RR∆S®MinoriMirari®.Prod@lemmy.worldBanned from community
    link
    fedilink
    English
    arrow-up
    2
    arrow-down
    3
    ·
    2 months ago

    current maintenance on the generator is on going. the actual token length is 512 which is roughly 512 itemizations not exactly words but if using natural language then it makes sense to say words. this doesn’t account for multi modality aka negatives and a few system to embed features “(())” are not counted in this, although right now those features are down.

    • anonfox421@lemmy.worldOP
      link
      fedilink
      English
      arrow-up
      1
      arrow-down
      1
      ·
      2 months ago

      Do you have a source you could link to for the ongoing maintenance note please?

      The main thing I’m curious about and hoping for an answer regarding is whether this is the expected behaviour following the update and something we should get used to, or should we expect it to return to how it was with longer prompts working?

      • RudBo@lemmy.world
        link
        fedilink
        English
        arrow-up
        3
        arrow-down
        1
        ·
        2 months ago

        MinoriMirari®.Prod is a well-known spammer and speak crazy codes. Not that everything he says must be incorrect but sure it’s always confusing.

        • RR∆S®MinoriMirari®.Prod@lemmy.worldBanned from community
          link
          fedilink
          English
          arrow-up
          1
          arrow-down
          2
          ·
          2 months ago

          Could have just directed him too Reddit. Anyway that is where you’d find source information. I’am just telling you what I knoe from observation.

          • anonfox421@lemmy.worldOP
            link
            fedilink
            English
            arrow-up
            1
            ·
            2 months ago

            Would you be so good as to link to the Reddit post you’re referring to, as I haven’t seen anything on the Reddit, Discord, or here about the maintenance still being ongoing and I’d be interested to see that post.

      • RR∆S®MinoriMirari®.Prod@lemmy.worldBanned from community
        link
        fedilink
        English
        arrow-up
        2
        arrow-down
        1
        ·
        2 months ago

        You typically count every word if not more then that into a single token less you use sectioning or have a lot of conjunctions in play. the model is the 512 model, so the onky way 300 words cut off makes sense is if Main Dev turned off half the systen considering it runs two independent 256 token blocks.

        • RR∆S®MinoriMirari®.Prod@lemmy.worldBanned from community
          link
          fedilink
          English
          arrow-up
          1
          arrow-down
          1
          ·
          2 months ago

          Although like I said, Its probably temporary mate. I assist in coding non natural language, algorithmic formula, and similar things while conceptually developing The T2iFramework/&plugin as assistant To the main Dev. No official final details are given other then we are running on base Schnell right now likely with only 1 token block running. Also, Multi Modalities are shut off which are responsible for a range of things.

  • MotionMuse@lemmy.world
    link
    fedilink
    English
    arrow-up
    1
    arrow-down
    4
    ·
    2 months ago

    Yeah the site got alot worse. We dont know if its gonna be like this from now or not.

    Dev, please give us some news announcement, anything will be appreaciated!

    I also am very disappointed in the update, it took away so much more than it gave, Why remove weights and reduce the token count to so small?

    And why the refusal to give simple communication about it?