import spaces # Must precede torch for ZeroGPU CUDA emulation. import math import os import gradio as gr import torch from geometry import output_size from pipeline import PRESETS, PRESET_FIELDS, run_pipeline, work_estimate MODELS = {} if os.environ.get('TILE_V3_UI_ONLY') != '1': from engine import RestorationModel torch.set_num_threads(2) MODELS['Restore'] = RestorationModel('Restore') MODELS['Fast'] = RestorationModel('Fast') # The multi-step teacher shares the resident VAE and vision encoder with Restore. MODELS['Detail'] = RestorationModel('Detail', shared=MODELS['Restore']) def duration(image, mode='Restore', size='2×', seed=42, restoration=1., color_strength=1., batch_size=4, passes=3, pass2_strength=.5, pass3_strength=.3, texture_only=True, detail_boost=.25, local_contrast=.1, overlap=128, variations=1, steps=10, guidance=.5, weak_condition=.1, preset='Maximum detail', progress=None, **kwargs): try: _, seconds = work_estimate(output_size(*image.size, size), mode, passes, overlap, variations, steps, batch_size) return min(180, max(20, math.ceil(seconds + 10))) except (AttributeError, ValueError, TypeError, ZeroDivisionError): return 20 @spaces.GPU(duration=duration) def upscale(image, mode='Restore', size='2×', seed=42, restoration=1., color_strength=1., batch_size=4, passes=3, pass2_strength=.5, pass3_strength=.3, texture_only=True, detail_boost=.25, local_contrast=.1, overlap=128, variations=1, steps=10, guidance=.5, weak_condition=.1, preset='Maximum detail', progress=gr.Progress()): try: return run_pipeline(MODELS, image, mode, size, seed, restoration, color_strength, batch_size, passes, pass2_strength, pass3_strength, texture_only, detail_boost, local_contrast, overlap, variations, steps, guidance, weak_condition, preset, progress) except ValueError as e: raise gr.Error(str(e)) from e except torch.cuda.OutOfMemoryError as e: torch.cuda.empty_cache() raise gr.Error('GPU memory exceeded. Reduce tile batch size, output size, or use Restore mode.') from e def select_variant(state, evt: gr.SelectData): if not state: return gr.skip(), gr.skip() index = evt.index[0] if isinstance(evt.index, (list, tuple)) else evt.index if not isinstance(index, int) or not 0 <= index < len(state['files']): return gr.skip(), gr.skip() path = state['files'][index] return (state['reference'], path), path CSS = ''' .gradio-container {max-width: 1280px !important;} #hero {padding: 12px 0 22px;} #hero h1 {font-size: clamp(2rem, 4vw, 3.25rem); letter-spacing: -.04em; margin-bottom: .4rem;} #hero p {font-size: 1.05rem; opacity: .8;} #run {min-height: 50px; font-size: 17px;} /* Gradio's first layer is in flex flow, while the second is absolute. Give both the same viewport so contain sizing and zoom stay registered. */ #comparison .content > img { position: absolute !important; inset: 0 !important; width: 100% !important; height: 100% !important; max-height: none !important; margin: 0 !important; object-fit: contain !important; object-position: center !important; transform-origin: top left !important; } footer {display: none !important;} ''' with gr.Blocks(title=' ', theme=gr.themes.Soft(primary_hue='indigo', neutral_hue='slate'), css=CSS, delete_cache=(3600, 3600)) as demo: with gr.Row(): with gr.Column(scale=4): image = gr.Image(type='pil', image_mode='RGBA', label='Your image', sources=['upload', 'clipboard'], height=300) preset = gr.Dropdown(list(PRESETS) + ['Custom'], value='Maximum detail', label='Pipeline preset', info='Presets set starting values. Every control remains adjustable.') mode = gr.Radio([('Restore', 'Restore'), ('Fast', 'Fast'), ('Detail · slower', 'Detail')], value='Restore', label='Model', info='Restore uses VOSR 2.0. Detail uses a multi-step first pass, then Restore for refinements.') size = gr.Dropdown(['2×', '4×', '2048 px long edge', '4096 px long edge'], value='2×', label='Output size') with gr.Row(): passes = gr.Radio([1, 2, 3], value=3, label='Restoration passes') variations = gr.Dropdown([1, 2, 3], value=1, label='Seed variants', info='Distinct alternatives; more variants take longer.') run = gr.Button('Create detailed image', variant='primary', elem_id='run') with gr.Column(scale=6): comparison = gr.ImageSlider(label='Before / after', type='pil', height=460, elem_id='comparison') with gr.Row(): download = gr.File(label='Selected full-resolution PNG') bundle = gr.File(label='All variants + settings ZIP') status = gr.Markdown('Maximum detail is ready: Restore → two gentle texture passes → mild finishing.') gallery = gr.Gallery(label='Choose your result', columns=3, height=260, object_fit='contain', preview=False) selected_state = gr.State() with gr.Accordion('Refinement & finishing', open=False): with gr.Row(): restoration = gr.Slider(0, 1, value=1, step=.05, label='First pass strength') pass2_strength = gr.Slider(0, 1, value=.5, step=.05, label='Second pass strength') pass3_strength = gr.Slider(0, 1, value=.3, step=.05, label='Third pass strength') with gr.Row(): texture_only = gr.Checkbox(value=True, label='Texture-only refinements', info='Later passes add small detail while retaining broader shapes.') color_strength = gr.Slider(0, 1, value=1., step=.05, label='Original color preservation', info='At 1, keeps original hue, saturation and broad brightness while retaining restored texture.') with gr.Row(): detail_boost = gr.Slider(0, 1, value=.25, step=.05, label='Fine-detail boost', info='Luminance sharpening with local halo limits; increases crispness.') local_contrast = gr.Slider(0, 1, value=.1, step=.05, label='Local contrast') with gr.Accordion('Sampling & tiles', open=False): with gr.Row(): seed = gr.Number(value=42, precision=0, minimum=0, maximum=2147483647, label='Seed') overlap = gr.Dropdown([64, 128, 192], value=128, label='Tile overlap (px)', info='More overlap can improve texture continuity but adds work.') batch_size = gr.Dropdown([1, 2, 4, 8], value=4, label='Tile batch size') with gr.Column(visible=False) as sampler: gr.Markdown('**Detail model only:** these controls affect the multi-step first pass. Subsequent passes use Restore.') with gr.Row(): steps = gr.Dropdown([5, 10, 15, 25], value=10, label='Sampling steps') guidance = gr.Slider(-.5, 2, value=.5, step=.1, label='Fidelity guidance', info='Higher favors the input; lower gives more generative freedom.') weak_condition = gr.Slider(.05, .25, value=.1, step=.01, label='Weak conditioning', info='Higher tends toward fidelity; lower allows freer textures.') gr.Examples(examples=[['image1.jpg'], ['image2.jpg'], ['image4.jpg'], ['image5.jpg']], inputs=[image], cache_examples=False) with gr.Accordion('About, models & attribution', open=False): gr.Markdown('VOSR by Rongyuan Wu and collaborators: **Fast** (0.5B one-step), **Restore** (VOSR 2.0 one-step), ' '**Detail** (1.4B multi-step). Shared spatial noise and Gaussian tile blending keep overlaps consistent. ' 'Generated textures can differ from the original.\n\n' '[VOSR code](https://github.com/cswry/VOSR) · [Weights and licenses](https://huggingface.co/CSWRY/VOSR)\n\n' 'Hosted interface by [Gökay Aydoğan](https://orcid.org/0000-0002-2343-9433). ' 'Please cite the hosted interface and upstream projects when used in academic work.\n\n' '```bibtex\n@software{aydogan2026tileupscalerv3,\n author = {Aydoğan, Gökay},\n' ' title = {Tile Upscaler v3},\n year = {2026},\n publisher = {Hugging Face},\n' ' url = {https://huggingface.co/spaces/gokaygokay/SD3-Long-Captioner}\n}\n```') preset_outputs = [mode, passes, restoration, pass2_strength, pass3_strength, texture_only, detail_boost, local_contrast, color_strength, overlap, variations, steps, guidance, weak_condition] def apply_preset(name): if name not in PRESETS: return [gr.skip()] * (len(PRESET_FIELDS) + 1) return list(PRESETS[name]) + [gr.update(visible=PRESETS[name][0] == 'Detail')] preset.change(apply_preset, inputs=preset, outputs=preset_outputs + [sampler], queue=False, api_name=False) mode.change(lambda value: gr.update(visible=value == 'Detail'), inputs=mode, outputs=sampler, queue=False, api_name=False) inputs = [image, mode, size, seed, restoration, color_strength, batch_size, passes, pass2_strength, pass3_strength, texture_only, detail_boost, local_contrast, overlap, variations, steps, guidance, weak_condition, preset] run.click(upscale, inputs=inputs, outputs=[comparison, download, bundle, gallery, status, selected_state], api_name='upscale', concurrency_limit=1, concurrency_id='restoration') gallery.select(select_variant, inputs=selected_state, outputs=[comparison, download], queue=False, api_name=False) if __name__ == '__main__': demo.queue(max_size=16, default_concurrency_limit=1).launch(show_error=True)