tristmeister/HEISS-UI v0.14.0

HEISS UI v0.14.0

on GitHub

Generated images in the feature mockups

Changed

• The existing animated feature mockups now show real generated images in
  their image areas, including the model thumbnail, render preview, phone
  gallery, search results, Hidden grid and upscale comparison.
• The website moved to heiss-ui.vercel.app, with
  new guides for every model family, popular graphics cards and Macs, and
  comparisons with other ComfyUI front ends. The Website link in Settings goes
  there, and old links forward.
• Pony V7 runs as its makers' own workflow does: its text encoder reads
  prompts padded to 768 tokens, and pictures start at 1280×1536.
• The Krea 2 enhancer runs at its own default strength of 1 instead of
  1.5, where its author says to start.
• The original Sana 1.6B models run at CFG 2, as NVIDIA's own ComfyUI
  workflows do. SANA 1.5 keeps 4.5.

Fixed

• Wan video comes out clean on a Mac. Wan 2.1 and Wan 2.2 5B start on the
  euler sampler there, since uni_pc corrupts video on Apple Silicon.
• A Mac gets files it can load. Where a text encoder comes in full
  precision too (T5-XXL for SD 3.5, UMT5 for Wan), a Mac downloads that one,
  since Macs can't load quantized fp8 files. The fp8 starter models say
  they're not for Macs and are never picked there.
• Speed models start at the step count they're made for. An
  SDXL-Lightning, Qwen-Image Lightning or MiniMax H3 Turbo file with its step
  count in the name (like 4step) runs at exactly that many steps. DMD2 starts
  at 4 steps, H3 Turbo at 8.
• MiniMax H3 clips start at 5 seconds (124 frames), the shortest length
  it's trained for, instead of 2.3 seconds.
• HunyuanVideo 1.5 runs with Tencent's settings. The shift follows the
  model (5 at 480p, 9 for 720p text to video, 7 for 720p image to video), the
  CFG-distilled models take the 50 steps they need for a clean result, and
  text to video starts at 121 frames. A 4-step lightx2v file runs as Fast.
• Anima Turbo runs as Turbo: 10 steps at CFG 1, as its card says, instead
  of 30 steps at CFG 4. Other Anima files start on er_sde, the card's sampler.
• HiDream no longer makes black images with Flux's CLIP-L. It takes only
  its own CLIP-L and CLIP-G, and offers their download when they're missing.
  Dev and Fast hide the negative prompt, which has no effect at their CFG 1.
• SDXL merges named "Pony v7" stay SDXL on a ComfyUI on another computer,
  where only the file name can tell. Only Pony V7 itself runs as AuraFlow.
• SD 2.x base models start at 512, the size they're made for.

Downloads

• Windows: heiss-ui-0.14.0-windows-x64.zip (Node.js included), then Start HEISS UI.bat
• macOS (Apple silicon): heiss-ui-0.14.0-macos-arm64.zip, then Start HEISS UI.command
• Linux: heiss-ui-0.14.0-linux-x64.zip, then Start HEISS UI.sh or npm start
• heiss-ui-0.14.0.zip is what installed copies download to update themselves. You don't need it.