README
Noct Q Studio
A Gradio demo for the exact Noct Q V4 INT8 convrot checkpoint. Includes text-to-image, single-image editing, five canvas presets, negative prompts, seed controls, PNG downloads and generation metadata embedded in PNGs.
Uses ComfyUI core because the published files are ComfyUI-format quantized weights, rather than a Diffusers pipeline. The checkpoint, companion weights and ComfyUI revision are pinned. No custom nodes or substitute model are used.
Default generation settings follow the model author's workflow: 25 steps, Euler sampler, simple scheduler, CFG 3. CFG 1 is faster and ignores the negative prompt. The supplied workflow is included as an upstream reference.
Deployment
Run python deploy.py after hf auth login. It creates a public Gradio Space
at <your-account>/noct-q-studio with ZeroGPU (zero-a10g). It will not
request a billed GPU or change an existing Space. Free ZeroGPU creation is
subject to Hugging Face account eligibility and the account's Space cap.
The first build downloads approximately 16 GB of model files. Model modules are loaded at startup on CPU; ComfyUI moves their native quantized tensors inside the actual GPU worker, avoiding fake-CUDA packing of tensor subclasses. GPU duration estimates are provisional until a live generation is measured. If the account cannot create ZeroGPU Spaces, deployment stops; it does not fall back to paid hardware or CPU inference.
For GPU hardware outside ZeroGPU, the platform's spaces decorator is a no-op.
Paid hardware is not selected by the included deployment script.
API
Gradio exposes /generate and an MCP server. The generate function accepts
prompt, negative_prompt, size, steps, guidance, seed,
randomize_seed, and an optional reference image. It returns a PNG preview,
PNG download, actual seed, and a generation-details string. PNG files are
temporary and are removed after 24 hours; download images you want to keep.
License
The model is a derivative of Qwen Image 2.1 under the Qwen Research License, for non-commercial use. See the upstream model's LICENSE and NOTICE: https://huggingface.co/Noctaluna/Noct-Q-Uncensored-Qwen-Image-2.1
ComfyUI is GPL-3.0; its unmodified pinned source is fetched from its upstream repository at startup. This demo's Python code is provided under GPL-3.0.
README
Noct Q Studio
A Gradio demo for the exact Noct Q V4 INT8 convrot checkpoint. Includes text-to-image, single-image editing, five canvas presets, negative prompts, seed controls, PNG downloads and generation metadata embedded in PNGs.
Uses ComfyUI core because the published files are ComfyUI-format quantized weights, rather than a Diffusers pipeline. The checkpoint, companion weights and ComfyUI revision are pinned. No custom nodes or substitute model are used.
Default generation settings follow the model author's workflow: 25 steps, Euler sampler, simple scheduler, CFG 3. CFG 1 is faster and ignores the negative prompt. The supplied workflow is included as an upstream reference.
Deployment
Run python deploy.py after hf auth login. It creates a public Gradio Space
at <your-account>/noct-q-studio with ZeroGPU (zero-a10g). It will not
request a billed GPU or change an existing Space. Free ZeroGPU creation is
subject to Hugging Face account eligibility and the account's Space cap.
The first build downloads approximately 16 GB of model files. Model modules are loaded at startup on CPU; ComfyUI moves their native quantized tensors inside the actual GPU worker, avoiding fake-CUDA packing of tensor subclasses. GPU duration estimates are provisional until a live generation is measured. If the account cannot create ZeroGPU Spaces, deployment stops; it does not fall back to paid hardware or CPU inference.
For GPU hardware outside ZeroGPU, the platform's spaces decorator is a no-op.
Paid hardware is not selected by the included deployment script.
API
Gradio exposes /generate and an MCP server. The generate function accepts
prompt, negative_prompt, size, steps, guidance, seed,
randomize_seed, and an optional reference image. It returns a PNG preview,
PNG download, actual seed, and a generation-details string. PNG files are
temporary and are removed after 24 hours; download images you want to keep.
License
The model is a derivative of Qwen Image 2.1 under the Qwen Research License, for non-commercial use. See the upstream model's LICENSE and NOTICE: https://huggingface.co/Noctaluna/Noct-Q-Uncensored-Qwen-Image-2.1
ComfyUI is GPL-3.0; its unmodified pinned source is fetched from its upstream repository at startup. This demo's Python code is provided under GPL-3.0.