CivArchive
    Rubberhose Ruckus HiDream - v1.0
    Preview 74618836Preview 74618842Preview 74618832Preview 74618829Preview 74618849Preview 74618834Preview 74618835Preview 74618850Preview 74618853Preview 74618828Preview 74618852Preview 74618854Preview 74618831Preview 74618839Preview 74618837Preview 74618825Preview 74618845Preview 74618846Preview 74618848Preview 74618851

    Rubberhose Ruckus HiDream


    Ruberhose Ruckus HiDream LoRA is LyCORIS-based and trained to replicate the iconic vintage rubber hose animation style of the 1920s–1930s. With bendy limbs, bold linework, expressive poses, and clean color fills, this LoRA excels at creating mascot-quality characters with a retro charm and modern clarity. It's ideal for illustration work, concept art, and creative training data. Expect characters full of motion, personality, and visual appeal.

    I recommend using the LCM sampler and Simple scheduler for best quality. Other samplers can work but may lose edge clarity or structure. The first image includes an embedded ComfyUI workflow — download it and drag it directly into your ComfyUI canvas before reporting issues. Please understand that due to time and resource constraints I can’t troubleshoot everyone's setup.

    Trigger Words: rubb3rh0se, mascot, rubberhose cartoon
    Recommended Sampler: LCM
    Recommended Scheduler: SIMPLE
    Recommended Strength: 0.5–0.6
    Recommended Shift: 4.0–5.0

    v1: Text appears when not prompted for, I included some images with text thinking I could get better font styles in outputs but it introduced overfitting on text.

    v2: Training for v2 includes some generations from the v1 model and more focus on variety. I feel it's a good balance between creativity and coherence. Both versions work well, v2 is more flexible but might need a little more prompting.

    Training ran for 2500 steps, 2 repeats at a learning rate of 2e-4 using Simple Tuner on the main branch. The dataset was composed of 96 curated synthetic 1:1 images at 1024x1024. All training was done on an RTX 4090 24GB, and it took roughly 3 hours. Captioning was handled using Joy Caption Batch with a 128-token limit.

    I trained this LoRA with Full using SimpleTuner and ran inference in ComfyUI with the Dev model, which is said to produce the most consistent results with HiDream LoRAs.

    If you enjoy the results or want to support further development, please consider contributing to my KoFi: https://ko-fi.com/renderartist
    renderartist.com

    Description

    LORA
    HiDream

    Details

    Downloads
    120
    Platform
    CivitAI
    Platform Status
    Available
    Created
    5/6/2025
    Updated
    9/27/2025
    Deleted
    -
    Trigger Words:
    rubb3rh0se

    Files

    Rubberhose_Ruckus_HiDream_v1_renderartist_2500.safetensors