Instructions to use PrunaAI/P-Image-OSS-Comfy with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusion Single File
How to use PrunaAI/P-Image-OSS-Comfy with Diffusion Single File:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Pruna AI
How to use PrunaAI/P-Image-OSS-Comfy with Pruna AI:
from pruna import PrunaModel model = PrunaModel.from_pretrained("PrunaAI/P-Image-OSS-Comfy") - Notebooks
- Google Colab
- Kaggle
Sporadic execution_error at PH3ImageConditioning (~18% of batch runs) with the int8 convrot build β measured stats
We run the ComfyUI build of P-H3-Image from this repo (full model, int8 convrot: UNET, Qwen3-VL 32B based text encoder, VAE) in ComfyUI 0.45.0 for batch generation with multi-image reference conditioning (up to 9 reference images, each scaled to 2.0 total megapixels before conditioning). Output is approx. 1184 x 1760. Hardware: RTX 5060 Ti, 16 GB VRAM, 64 GB system RAM, Windows. Jobs are submitted via the API by a driver script, one at a time.
Measured across four batches on 2026-11-14/15:
- 408 jobs attempted, 73 failed (about 18%)
- 100% of the failures were an
execution_errorat thePH3ImageConditioningnode; no failures at any other node - The failures are not deterministic: re-running a failed job with the identical prompt, seed (e.g. 20262010) and reference set succeeded on the first retry
- Most failures abort after roughly 30 seconds, while successful jobs take about 230 seconds. During one window with two concurrent API submitters, 12 jobs instead ran 250-260 seconds before failing at the same node
The full report, including per-batch statistics, is filed with ComfyUI here: https://github.com/Comfy-Org/ComfyUI/issues/17807
We are cross-posting here because of a hypothesis we cannot verify on our own (explicitly unverified): even in its int8 build, the text encoder cannot reside fully in 16 GB of VRAM, so ComfyUI offloads part of it to system RAM, and we suspect the failures correlate with memory pressure during conditioning β the worsening under two concurrent submitters is consistent with that. Our driver truncated the error detail string, so we also do not have the full exception text (partial strings looked like "Fault failed: 2").
Questions for the model/build side:
- Is this failure mode known for the int8 convrot build, in particular when the text encoder is offloaded to system RAM?
- Are there recommended minimum VRAM figures or settings for the conditioning step on 16 GB cards?
- What additional data would help β VRAM/RAM traces at failure time, full server logs, the exact workflow JSON? We can supply the workflow and per-job logs (timestamps, durations, seeds, success/failure) for all 408 runs.