Skip to content

Tags

Tags give the ability to mark specific points in history as being important
  • v0.9.0

  • v0.7.0

    v0.7.0: Qwen-Image-Edit + Sana 1.5, FLUX encoder validation (black-image fix)
  • v0.6.0

    emoji signature in filenames + PNG metadata
  • v0.5.0

    img2img (web UI) + Docker deployment
  • v0.4.0

    v0.4.0 — web UI with real-time generation
  • v0.3.1

    v0.3.1 — FLUX Dev + Krea downloaded, download_models.sh rewrite
    
    All three FLUX models operational and tested (seed 42, same prompt):
      Schnell  4 steps  — fast, cinematic, sharp
      Dev     20 steps  — photorealistic, detailed skin, dramatic light
      Krea    25 steps  — natural/film look, no AI artifacts, closest to real photo
    
    download_models.sh now handles all models in one script with HF_TOKEN support.
  • v0.3.0

    v0.3.0 — FLUX.1/2 generation support
    
    New architecture: FLUX (flow-matching transformer)
      Auto-detected by checkpoint name — no flags needed
      Separate workflow: UNETLoader + DualCLIPLoader (T5+CLIP) + FluxGuidance
      No negative prompt (FLUX architecture)
      Quality presets tuned for FLUX (guidance 0–4, euler/simple/beta)
      Native FLUX resolutions (832×1216, 1344×768, 1024×1024…)
    
    Models installed:
      flux1-schnell-fp8  16.4 GB  4-step, Apache 2.0, best speed/quality
      t5xxl_fp8           4.6 GB  shared T5-XXL text encoder
      clip_l              234 MB  shared CLIP-L encoder
      flux-ae             159 MB  FLUX VAE
    
    Ready to use (pending HF_TOKEN):
      flux1-dev          — 20-step high quality
      flux1-krea-dev     — natural/less AI-looking style
      flux2-dev          — maximum quality from BFL
    
    Stack: PyTorch 2.11+cu130, ComfyUI, aarch64, 120 GB VRAM
  • v0.2.0

    v0.2.0 — quality upgrade: SDXL, FreeU, ESRGAN, VAE
    
    Image generation improvements:
      - Juggernaut XL v9: best photorealism model, no hallucinations
      - Realistic Vision V6: fast SD 1.5 photorealism fine-tune
      - SDXL auto-detection by checkpoint name — correct resolutions/CFG/
        scheduler applied automatically
      - --quality fast|normal|high|ultra presets (separate tables for
        SD 1.5 and SDXL)
      - FreeU v2 enabled by default (free detail boost)
      - External VAE support (vae-ft-mse-840000 for better SD 1.5 colours)
      - ESRGAN pixel upscaler in hi-res fix (4x-UltraSharp, sharper than bislerp)
      - CLIP skip control
      - --list-vae / --list-upscalers commands
    
    Models added to ComfyUI:
      checkpoints/Juggernaut-XL-v9.safetensors      (6.7 GB, SDXL)
      checkpoints/Realistic_Vision_V6.safetensors   (4.0 GB, SD 1.5)
      upscale_models/4x-UltraSharp.pth              (64 MB)
      vae/vae-ft-mse-840000.safetensors             (320 MB)
  • v0.1.0

    v0.1.0 — initial release
    
    Core features:
      - txt2img: 17 resolution presets (256px → 4K), hi-res fix, batch,
        prompt from string or file (-f), all ComfyUI samplers/schedulers
      - img2vid: AnimateDiff, context window for long sequences, 2K upscale,
        prompt from string or file (-f)
      - ./kino CLI: unified entry point for all tools
      - pixel_art: 8 retro palettes, GIF export, CRT/scanline effects
      - ascii: video/image → ASCII, realtime audio spectrum visualiser
      - manim: math & algorithm animations (Fourier, sort, BFS, Euler)
      - p5.js: 7 interactive browser demos + GLSL shaders
      - suno: Suno AI music client with 6 style presets
      - mcp_server: Claude Desktop integration
    
    Stack: Python 3.12, PyTorch 2.11+cu130, ComfyUI, AnimateDiff-Evolved,
           aarch64, NVIDIA GB10 120 GB VRAM, CUDA 13.0