You can deploy this model and start earning money today!

xwk123/
temp

Views

No views yet

History-W3 HTML World Model

This upload contains the deployable Qwen3-VL-8B full-finetuned world model used to predict the next mobile GUI screen as complete HTML.

Model

Hub path: html_world_model/

The model input contains the current screenshot, up to three historical screenshots, the complete previous action history, and the current action. The output is the complete HTML for the predicted next screen.

Only the final deployable Hugging Face model files are uploaded. Training-time checkpoint-*, optimizer, DeepSpeed, and RNG state directories are intentionally excluded.

Inference code

inference/eval_world_model_history_cmgui.py: history-w3 HTML inference.
inference/render_cmgui_html_outputs.py: render predicted HTML with Chromium/Playwright.
inference/prepare_worldmodel_regions.py: detect image slots and create green repaint conditions.
inference/infer_worldmodel_regions.py: run the GUIdiff LoRA and composite repainted regions.
inference/visualize_cmgui_pipeline.py: create pipeline comparison sheets.
training/run_world_model_direct_full_html_history_w3_16gpu_pai.sh: original 16-GPU training launcher.

The HTML inference script expects an OpenAI-compatible model server and CMGUI pair/episode metadata. See each script's --help output for the exact arguments.

xwk123/temp

History-W3 HTML World Model

Model

Inference code

xwk123/
temp