diff options
| author | Danilo M. <danix@danix.xyz> | 2026-07-08 16:37:52 +0200 |
|---|---|---|
| committer | Danilo M. <danix@danix.xyz> | 2026-07-08 16:37:52 +0200 |
| commit | 2a80aea10ce31c9081692c4d516044a29e39a0c5 (patch) | |
| tree | 0581049588ce229fc02488f66a2f6c2e8d8c8c4e /generate.py | |
| parent | 1ba3a53299bc9c3e8781a9ca99338f4c20915506 (diff) | |
| download | imggen-2a80aea10ce31c9081692c4d516044a29e39a0c5.tar.gz imggen-2a80aea10ce31c9081692c4d516044a29e39a0c5.zip | |
Fix pony repo and surface daemon swap failures (Arc integration)
Integration on the real Arc B580 uncovered two issues:
- The pony repo AstraliteHeart/pony-diffusion-v6-xl is gated (HF returns
401 Repository Not Found), so swaps to it failed. Point the pony key at
kitty7779/ponyDiffusionV6XL, an ungated diffusers-format mirror with
safetensors weights that loads cleanly via the existing fp16 fallback.
- The wrapper swallowed a failed model swap: under set -e the unguarded
swap curl aborted the script with exit 0 and no message. Capture the
/model response and HTTP status, print the daemon's error, and exit 1
when the swap does not return 200.
sdxl, realvis, and pony all verified end to end (one-shot + daemon),
output owned by the host user, peak within the 12GB budget.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Diffstat (limited to 'generate.py')
| -rw-r--r-- | generate.py | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/generate.py b/generate.py index 9cbc167..95b2dd7 100644 --- a/generate.py +++ b/generate.py @@ -27,7 +27,7 @@ from diffusers import StableDiffusionXLPipeline, AutoencoderKL MODELS = { "sdxl": "stabilityai/stable-diffusion-xl-base-1.0", # default, general "realvis": "SG161222/RealVisXL_V5.0", # photoreal, permissive - "pony": "AstraliteHeart/pony-diffusion-v6-xl", # unrestricted, versatile + "pony": "kitty7779/ponyDiffusionV6XL", # unrestricted, versatile (ungated diffusers-format mirror) } DEFAULT_MODEL = "sdxl" # fp16-fix VAE avoids the fp32 upcast that OOMs a 12GB card at decode time. |