GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Timetocreate7777
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Timetocreate7777
Model Tree
unsloth/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
The adapter was trained to recognize and return Director AI camera coordinates consisting of:
The associated ComfyUI custom node converts these coordinates into deterministic, tested instructions for the required image-generation workflow. Exact prompt corrections, profile enforcement, output naming, and workflow metadata are handled by code rather than relying on the language model to calculate them.
Director_AI_Qwen_image_control.json, included in this repositoryThis project uses two different LoRAs with separate responsibilities:
The adapter in this repository runs with the unsloth/Qwen3.5-4B base model. It interprets natural requests and returns P–E–D coordinates.
The image-generation workflow requires the separate Qwen Image Edit 2511 Multiple-Angles LoRA used during development and testing. It performs the actual camera-view transformation.
The ordinary Qwen Image Edit model without the required Multiple-Angles LoRA was not used to validate these controls and should not be treated as an equivalent replacement.
Third-party base-model and image-LoRA weights are not included in this repository. Download them from their official sources and follow their respective licenses.
| Code | Camera position |
|---|---|
| P1 | Front |
| P2 | Front Right |
| P3 | Right |
| P4 | Back Right |
| P5 | Back |
| P6 | Back Left |
| P7 | Left |
| P8 | Front Left |
| Code | Camera elevation |
|---|---|
| E0 | Extreme Low Angle |
| E1 | Low Angle |
| E2 | Eye Level |
| E3 | Elevated |
| E4 | High Angle |
| E5 | Extreme Elevated View |
E5 is intentionally described as Extreme Elevated View, not a guaranteed 90-degree directly overhead view. Output testing showed that the required image LoRA does not consistently maintain a true vertical overhead position.
| Code | Framing distance |
|---|---|
| D1 | Close |
| D2 | Medium |
| D3 | Wide |
Example: P3-E3-D2 requests a medium-distance elevated view from the subject's right side.
This adapter is intended to be loaded by the Director AI ComfyUI custom node for natural-language camera interpretation.
Dropdown mode does not require the interpreter model: P, E, and D selections are compiled deterministically by the workflow. The LoRA is used when the user selects natural-language mode.
Download the complete repository, then run install_windows.bat.
Do not download only the workflow JSON. The workflow requires the included
ComfyUI-Director-AI-PED custom-node package.
Use the included Director AI installer, custom node, and:
text
Director_AI_Qwen_image_control.json
The adapter folder must contain at least:
text
adapter_config.jsonadapter_model.safetensorstokenizer.jsontokenizer_config.jsonchat_template.jinjaprocessor_config.json
Load the adapter with:
text
unsloth/Qwen3.5-4B
Do not rename third-party model files unless the installation documentation specifically requires it. Complete ComfyUI installation instructions and model placement information are supplied with the release package.
The final E0/E5 upgrade dataset contained:
The dataset retained 3,126 corrected examples from the previously successful training and added balanced coverage for P1–P8, E0–E5, and D1–D3. Examples included coordinate requests, natural-language descriptions, and structured response metadata.
The image model itself was not trained as part of this LoRA. This adapter trains only the Qwen3.5-4B language interpreter.
Observed final training loss was approximately 0.0343, while evaluation loss stabilized near 0.030.
The held-out evaluation set contained 240 cases covering P1–P8, E0/E5 upgrade behavior, D1–D3, natural-language requests, and associated workflow metadata.
| Measurement | Result |
|---|---|
| Strict full-response score | 205/240 — 85.42% |
| Correct P/E/D coordinates | 238/240 — 99.17% |
| E-coordinate accuracy | 240/240 — 100% |
| D-coordinate accuracy | 240/240 — 100% |
| P-coordinate errors | 2 |
| Metadata-only errors | 33 |
Thirty metadata errors involved incorrect rotation-degree fields and three involved safe-crop flags. These metadata fields are calculated deterministically by the production workflow and are not trusted to the language model. Only two evaluation cases produced incorrect camera coordinates; both involved natural-language right-side interpretation.
Dropdown mode is deterministic and does not depend on language-model interpretation.
Director AI is intended to reduce wasted computation by making usable camera prompts easier to produce and reducing repeated failed generations.
Director AI is free to try. Optional support can help fund:
markdown
SUPPORT OR DONATION LINK: [[ADD LINK OR REMOVE THIS SECTION](https://www.paypal.com/ncp/payment/3JR9PABGZC4GA)]
When reporting a problem, please include:
This project builds on Qwen3.5, Unsloth, Hugging Face Transformers and PEFT, ComfyUI, Qwen Image Edit, and the third-party Multiple-Angles image LoRA. Their respective trademarks, model weights, and licenses remain with their creators.