residual_fusion

ResidualFusion Prompt Controller

ResidualFusion Prompt Controller is a supervised BF16 LoRA adaptation of Qwen3-1.7B. It converts conversational photo-editing requests into the constrained control schema consumed by the ResidualFusion image pipeline.

The model does not generate or modify pixels. Image analysis, residual parameter prediction, semantic masks, safety selection, and rendering remain deterministic components of the ResidualFusion application.

Model Assets

SHA-256 hashes and exact file sizes are recorded in the v1.0.0 asset manifest. The complete editor is published in the AI Photo Editor repository.

Training

• Base model: Qwen/Qwen3-1.7B, frozen during adapter training.
• Method: supervised BF16 LoRA with rank 16, alpha 32, and dropout 0.05.
• Selected checkpoint: step 260, chosen by intent accuracy, complete-field exact accuracy, and then the earliest checkpoint.
• Output contract: structured intent, strength, and preservation constraints for the downstream image editor.

Evaluation

The selected validation checkpoint achieved:

• Intent accuracy: 100%.
• Complete-field exact accuracy: 95%.
• Constraint micro-F1: 98.46% on the selected validation split.

An isolated synthetic prompt audit achieved 97.5% intent accuracy, while a harder compound-constraint audit achieved 74.03% micro-F1. These are task-specific control metrics and should not be interpreted as general language-model accuracy.

Ollama Deployment

Use the installer included in the ResidualFusion application package:

.\install_prompt_model.ps1 -RepoId "Kaiii1912/residual_fusion"
.\install_prompt_model.ps1 -CheckOnly

The installer downloads the GGUF, verifies its SHA-256 hash, and creates the local Ollama model ai-photo-prompt-control:exp007-v1.

Limitations

• The model is specialized for ResidualFusion's constrained editing schema; it is not intended as a general chatbot.
• Compound constraints remain less reliable than single editing intents.
• Reported results are validation and isolated audit measurements. A new, untouched natural-language blind test has not been opened.
• The complete photo-editing result also depends on the separate ResidualFusion image pipeline and cannot be reproduced from this language model alone.

License

The base Qwen3-1.7B model is distributed under Apache-2.0. Use of the application code and other third-party components remains subject to their respective licenses.