v0.3.0

New

• MOSS-Image. OpenMOSS's 5B model, Base and SFT, with two Get models cards for the
  SFT: Comfy-Org's official INT8 file (5.4 GB, which sd.cpp runs as INT8) and the BF16
  original (10.2 GB). The Convert tab lists an INT8 file but won't convert it, since
  its weights are stored rotated. It renders 50 steps at CFG 3, as its model card does. It needs a fourth
  file beside the model, encoder and VAE: Qwen3.5's tokenizer.json, which the card
  fetches and the Models panel lists as Tokenizer. Its Qwen3.5-9B encoder is never
  offered to Flux.2 Klein 9B, or Klein's Qwen3-8B to it, although both are 4096 wide.
• Get models is grouped: image models and upscalers (and video models, with the
  nightly Video tab on) each under a heading, in the list and in its menu.
• Qwen-Image 2.1. Qwen's 7B model, with its own VAE and Qwen3-VL-8B as the
  encoder, and a Get models card that also fetches Comfy-Org's turbo LoRA: with
  Turbo on it renders 8 steps at CFG 1 on the model's fixed turbo schedule,
  instead of 40 at CFG 6. Its turbo LoRA is recognised as a Qwen-Image 2.1 one, not a
  Qwen-Image one. Read its licence (Qwen's research licence) before you use it.
• Iris-3B and Iris-3B Turbo. Sperid Labs' small model, which draws in pixels with
  no VAE: just the model and Krea 2's Qwen3-VL-4B encoder, loaded as Q8 (about 8 GB on
  the card). The Turbo card is the one to start with — 4 steps at CFG 1 on its
  published schedule, where the base runs 100 at CFG 3 (200 network passes a
  picture). The Turbo can't be told from the base by its contents, so the downloaded
  file brings its settings with it.
• Layout boxes for MOSS-Image. Draw boxes on the picture to say where things go
  and what they are (an object, a character, or text with its exact words); the Prompt
  describes the whole scene. A Picture | Layout switch above the picture brings the
  boxes back after a render. Every pass is told the boxes again at its own size, and
  they're saved with the work item and in the PNG.
• Turbo for MOSS-Image. The MOSS cards also fetch lodestones' 8-step DMD LoRA,
  and a Turbo switch in the Models panel turns it on with its 8 steps at CFG 1
  (on by default, remembered per model), several times faster than 50 steps at CFG 3.
  Off puts the model's own steps and CFG back. MOSS LoRAs are now recognised as MOSS
  ones, so they're no longer offered to other models.
• Recommended models, for your GPU. FLUX.3 Klein, MOSS-Image and Krea 2 are
  marked Recommended on their Get models cards and menu buttons and in the Image
  tab's model list — where they fit the GPU's memory, which yasdig reads on Linux. On a
  card under 12 GB, Iris-3B Turbo is the recommendation instead. A card that needs more than
  the GPU has says so.
• Video tab (nightly — Settings → Nightly features; hidden otherwise):
  • Video models in Get models: Wan 2.2 TI2V-5B, the Wan 2.2 I2V-A14B pair, and
    MiniMax H3.
  • MiniMax H3 runs, with its own VAE (it was handed Wan's) and its sound. Its
    recipe is sd.cpp's (CFG 1), and it streams from the SSD because its 18 GB text
    encoder doesn't fit in System RAM's share. Check its licence before you use it.
  • Pick a model's parts (VAE, text encoder, H3's audio VAE) like on the Image
    tab. A reopened clip puts back the parts it was made with.
  • Clips are saved as a video file (.avi, with sound when the model makes it)
    in the output folder, beside the frames. Settings → Save clips as switches back
    to frames only. The player has a Download button and plays H3's sound, and
    Enlarge follows the same setting.
  • An Info panel, like the Image tab's: the clip's size, frames, seed (click
    to reuse it), how long it took, its settings, the files it ran with and where
    it was saved.
  • VAE tiling is on by default, since a clip is likelier to run out of memory in
    the VAE decode than anywhere else. Turn it off under Advanced for a faster decode
    when the card has room.

Full changelog: https://github.com/Bastardly/yasdig/blob/master/CHANGELOG.md

VirusTotal

Every release is scanned by VirusTotal once it is published. Each report shows what every engine said, and stays live: engines update their verdicts over time.

File	Result	Report
yasdig-windows-x86_64.zip	0 of 65 engines flagged it	report
yasdig.exe (inside the archive)	0 of 69 engines flagged it	report
yasdig-linux-x86_64.tar.gz	0 of 62 engines flagged it	report
yasdig (inside the archive)	0 of 63 engines flagged it	report

SHA-256

5782bd22e404ef7ce2035fbc1dae01c554e19c825dfd661e379d77ee7373a369  yasdig-windows-x86_64.zip
6b4e0d84bb73c7e3517cdc800f902df86dab3326551750116260e18e4135a87f  yasdig-linux-x86_64.tar.gz

Check a download with sha256sum -c SHA256SUMS (Linux) or certutil -hashfile <file> SHA256 (Windows), and where it was built with gh attestation verify <file> --repo Bastardly/yasdig.