<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/659f0806-a86e-47bd-9ea3-967a94209175/original=true/659f0806-a86e-47bd-9ea3-967a94209175.jpeg" />Installation — Folder StructureNightrun expects to live two levels inside your WebUI folder:<pre><code>YOUR-WEBUI-FOLDER\
 workflows\
 nightrun_2.0\ ← here</code></pre>A local Forge companion that builds structured prompt batches, runs them overnight, manages metadata-based upscales, and watches the boring parts, so your PC can do the babysitting for once.GitHub Link: <a target="_blank" rel="ugc" href="https://github.com/Bluenochian/Nightrun">https://github.com/Bluenochian/Nightrun</a><h1 id="nightrun-2.0">Nightrun 2.0</h1>A portable workflow system for Stable Diffusion Forge / Forge-Neo.Nightrun is not a model. Not a LoRA. Not a prompt file.It is the system that sits beside Forge and turns a messy local generation setup into something that can run for hours — without you checking whether the prompt broke, a LoRA name was wrong, the GPU is cooking itself, or the batch slowly turned into repetitive garbage.Nightrun is built for long local runs, editable prompt pools, metadata replay, LoRA-aware batching, upscale workflows, cooldown logic, and controlled randomness.Set the run. Leave it alone. Come back to something organized, replayable, and worth sorting through.<hr /><h2 id="what-nightrun-does">What Nightrun Does</h2>Nightrun builds prompts from editable pools, sends jobs through Forge, writes replayable metadata, manages upscales, scans LoRAs, protects long runs with cooldown logic, and keeps the whole thing configurable.It is made for people who want big batches without babysitting every prompt, every upscale, every progress bar, and every local setup problem.Controlled chaos. Not prompt soup.<hr /><h2 id="what's-new-in-2.0">What's New in 2.0</h2><h3 id="modular-architecture">Modular Architecture</h3>The engine was split from a single ~2055-line file into a purpose-driven module system:<ul><li><code>nightrun_runtime.py</code> — core shared logic</li><li><code>prompt_engine.py</code> — pool loading, weighted selection, prompt composition</li><li><code>lora_resolver.py</code> — LoRA scanning, aliases, tag construction</li><li><code>forge_client.py</code> — Forge API, txt2img/img2img, GPU guards</li><li><code>runner.py</code> — execution modes, output folders, cleanup, cooldowns, summaries</li><li><code>validation.py</code> — pool validation and dry-run reporting</li><li><code>nightrun_engine.py</code> — compatibility shim for older imports</li></ul><code>nightrun.py</code> is now a thin entrypoint that only imports and runs <code>main()</code>.<hr /><h3 id="three-engine-prompt-routing">Three-Engine Prompt Routing</h3>Prompt generation now routes through three distinct builders, selected by weighted random choice (default 1/1/1). You can control routing via <code>prompt_engines</code> in config, or disable it to revert to legacy behavior.Classic — Standard pool sampling with optional world module support. Closest behavior to 1.5, cleaned and optimized.Combinator — Builds on Classic output by applying controlled transformations:<ul><li>Internal combinators modify existing elements (e.g. a katana becomes a burning variant)</li><li>External combinators inject additional motifs (e.g. surreal, cursed, steampunk layers)</li><li>At least one combinator fires per prompt, capped at two</li></ul>Franchise — Forces generation through a world module from <code>franchise_catalog.py</code>. All visual elements are sourced from that universe for full coherence. Falls back to Classic if no module is available.<hr /><h3 id="prompt-cleanup-pipeline">Prompt Cleanup Pipeline</h3>A post-processing stage via <code>prompt_cleanup</code> now runs after every engine:<ul><li>Duplicate comma-tag removal</li><li>Concept-family repetition control</li><li>Compact SD-style prompt length enforcement</li><li>Removal of redundant subject labels when already implied</li></ul>Consistent prompt structure across all three engines.<hr /><h3 id="prompt-pools-full-rebuild">Prompt Pools — Full Rebuild</h3>All pools were rebuilt for SD-style tag efficiency. Verbose prose entries are gone. 2.0 uses compact, composable tags throughout.Pool 1.5 lines 2.0 lines concept.txt 3,601 154 genre.txt 8,317 2,743 glamour_pose.txt 4,241 761 clothing.txt 1,801 721 character.txt 701 44 body_focus.txt 1,105 481 sensual_layer.txt 1,921 529 hair.txt 2,401 790Removed: <code>character_loras.txt</code>, <code>style_loras.txt</code>, <code>lora_sampler_speed.txt</code>Added: <code>genre_core.txt</code> (69 lines), <code>genre_root.txt</code> (475 lines)A curated weighting system (<code>curated_top_count</code> / <code>curated_top_weight</code>) boosts high-quality early entries without removing full pool diversity.<hr /><h3 id="franchise-catalog-expanded-to-~50-modules">Franchise Catalog — Expanded to ~50 Modules</h3>Modules now use full dictionary-based definitions instead of compact tuples. The catalog grew from ~24 to ~50 entries:<code>alien</code> <code>batman</code> <code>bloodborne</code> <code>cowboy_bebop</code> <code>doctor_who</code> <code>dungeons_and_dragons</code> <code>elder_scrolls</code> <code>evangelion</code> <code>final_fantasy_vii</code> <code>game_of_thrones</code> <code>halloween</code> <code>hellraiser</code> <code>lord_of_the_rings</code> <code>matrix</code> <code>midsommar</code> <code>nier_automata</code> <code>nightmare_on_elm_street</code> <code>persona</code> <code>pokemon</code> <code>scream</code> <code>shadowrun</code> <code>terminator</code> <code>the_conjuring</code> <code>the_ring</code> <code>the_witcher</code> <code>vampire_masquerade</code> <code>x_men</code> <code>zelda</code>New filtering controls without touching the catalog directly: <code>enabled_module_ids</code> / <code>disabled_module_ids</code> / <code>enabled_source_names</code> / <code>disabled_source_names</code> / <code>label_allow_patterns</code> / <code>label_block_patterns</code><hr /><h3 id="selection-memory-retuned">Selection Memory — Retuned</h3><ul><li>Larger recent history windows across categories</li><li>Reduced repetition penalties for more natural variation</li><li>Added concept-level tracking via <code>recent_concepts</code></li><li>Dedicated repetition control for candle/light motifs</li><li>Curated weighting for high-quality biasing</li></ul><hr /><h3 id="config-changes">Config Changes</h3><code>fixed_suffix</code> updated for sharper SD output:<ul><li>Before: <code>clean coherent design, sharp focus, high detail, polished lighting</code></li><li>After: <code>crisp linework, sharp focus, high detail, clear material texture, clean color separation, high micro-contrast, dramatic lighting</code></li><li><code>base.steps_min</code> raised: 30 → 50</li><li>Output folders now use <code>nightrun-YYYYMMDD-HHMMSS</code></li><li><code>--reset-local</code> now preserves local tuning instead of overwriting it</li></ul><hr /><h3 id="new-vram-profile-tuned_6gb">New VRAM Profile — <code>tuned_6gb</code></h3>Setting Value Base resolution 1344×768 Upscale resolution 3072×1728 Steps 65 LoRA cap 7 VRAM threshold ≤ 6144 MB<hr /><h3 id="nsfw-sensual_layer-expansion">NSFW — sensual_layer Expansion</h3>The <code>sensual_layer</code> pool was fully reworked to properly support explicit NSFW generation paths. It now includes dedicated nudity, erotic styling, body exposure, lingerie, fetish-adjacent aesthetic, and mature composition tags designed to integrate cleanly with the SD-style prompt system and franchise/genre routing.NSFW is enabled by default. For SFW-only runs, open <code>config.json</code>, find <code>sensual_layer</code>, and set <code>"chance": 0</code>.<hr /><h2 id="core-features">Core Features</h2><h3 id="editable-prompt-pools">Editable Prompt Pools</h3>Nightrun uses plain <code>.txt</code> prompt pools instead of one giant hardcoded prompt. Edit pool files to change how the whole machine behaves:<code>concept</code> <code>genre</code> <code>genre_core</code> <code>genre_root</code> <code>character</code> <code>hair</code> <code>eyes</code> <code>clothing</code> <code>pose</code> <code>glamour_pose</code> <code>expression</code> <code>sensual_layer</code> <code>body_focus</code> <code>equipment</code> <code>environment</code> <code>lighting</code> <code>camera</code> <code>effects</code> <code>style_layer</code> <code>color_grade</code> <code>composition</code> <code>detail</code> and more.The world module system is deeper and optional. The pool system is where most users should start.<hr /><h3 id="controlled-randomness">Controlled Randomness</h3>Nightrun is random — but not stupid-random. You can configure:<ul><li>Pool chances and pick counts</li><li>Boosted chances when related pools appear</li><li>Base step ranges</li><li>LoRA chances, weight ranges, and caps</li><li>World module and fusion mode chances</li><li>Cooldown timing and upscale behavior</li><li>Selection memory and color repetition penalties</li></ul>You are not stuck with one fixed behavior. Tune the machine until it behaves like your machine.<hr /><h3 id="selection-memory">Selection Memory</h3>During a run, Nightrun remembers recent choices and penalizes repetition. It tracks broad color families — blue, cyan/teal, red, purple, green, gold, black, silver, iridescent — so if a batch starts drifting into one color mood, it pushes away.This is one of the main reasons long runs do not feel repetitive.<hr /><h3 id="rare-detail-control">Rare Detail Control</h3>Some details are only interesting when they stay rare. Heterochromia, for example, can be configured so it does not take over an entire batch. Special things stop feeling special when every third image has them.<hr /><h3 id="lora-handling">LoRA Handling</h3>Nightrun does not blindly slap LoRAs into the prompt. It can:<ul><li>Scan local LoRA folders and resolve names</li><li>Use aliases and soft-match ugly filenames</li><li>Skip missing LoRAs without crashing the run</li><li>Randomize LoRA weights within configured ranges</li><li>Add synergy/support LoRAs with controlled chances</li><li>Cap random and total LoRA counts</li><li>Deduplicate final LoRA tags</li></ul>Local LoRA libraries are messy. Civitai names, trigger names, and local filenames are often all slightly different. Nightrun tries to survive that.<hr /><h3 id="coherent-world-modules">Coherent World Modules</h3>The normal prompt pools are the everyday engine. The world module system is the coherence layer.When it activates, Nightrun selects a world/franchise module and pulls related elements from the same source family — character archetypes, outfits, environments, props, factions, technology, magic systems, lighting mood, color palettes, cinematic language, and more.The goal is not to replace the normal pools. The goal is to give the run a coherent themed spine when that system activates.Fusion mode allows controlled cross-world combinations while reducing risky generic pool chances — unusual mixes without the machine blindly dumping everything into one prompt.<hr /><h3 id="long-run-safety">Long-Run Safety</h3>GPU temperature pause — Nightrun watches GPU temperature through <code>nvidia-smi</code>. If the GPU gets too hot, it pauses. When it cools below the configured resume threshold, it continues.VRAM chill sessions — For big batches, Nightrun schedules random cooldown sessions. Run several generations, pause for a few minutes, let the system settle, continue. Small thing. For overnight workflows it matters.Between-image rest — Configurable sleep time between images so the machine is not being tortured nonstop.Nightrun is not just "go fast." It is "keep going safely."<hr /><h3 id="upscale-workflow">Upscale Workflow</h3>Nightrun supports both scheduled upscale workflows and metadata-based re-upscaling. It can:<ul><li>Generate base images and upscale during full runs</li><li>Auto-upscale low-step base renders</li><li>Accept drag-and-drop metadata for upscaling</li><li>Recreate from metadata if needed</li><li>Write new metadata beside upscaled images</li><li>Chain later upscales from previous upscale metadata</li></ul>A fast base image does not have to stay a rough base image. Low-step renders can go through upscale logic and recover quality without burning time on every single base generation.<hr /><h3 id="metadata-replay">Metadata Replay</h3>Every image gets a metadata sidecar with everything important saved:<code>seed</code> <code>prompt</code> <code>negative prompt</code> <code>selected pools</code> <code>selected modules</code> <code>LoRAs</code> <code>LoRA sources</code> <code>base steps</code> <code>status</code> <code>upscale reason</code> <code>errors</code> <code>GPU info</code>So later, you are not staring at a good image wondering what made it. The answer is saved beside it.<hr /><h3 id="validation-and-dry-run">Validation and Dry-Run</h3>Validation mode checks pool loading, prompt formatting, LoRA resolution, missing resources, alias behavior, and generation logic — before wasting GPU time.Dry-run mode writes prompt and metadata without calling Forge image generation. See what the run would do before committing to a real batch.Measure twice. Render once.<hr /><h3 id="prompt-variety-analyzer">Prompt Variety Analyzer</h3>A built-in tool scans old metadata and reports:<ul><li>Repeated categories, colors, and modules</li><li>Repeated prompt fragments</li><li>World module usage patterns</li><li>Exact prompt repeats</li><li>Signs that a batch is becoming too samey</li></ul>Tune the machine instead of guessing.<hr /><h2 id="outputs">Outputs</h2>Each run creates an organized folder:<pre><code>nightrun-YYYYMMDD-HHMMSS/
 0001_base_WIDTHxHEIGHT.png
 0001_upscale_WIDTHxHEIGHT.png
 0001_prompt.txt
 0001_metadata.json
 summary.json
 run.log
 contact_sheet.jpg
 validation_report.json
</code></pre>Exact files depend on the run mode.<hr /><h2 id="basic-usage">Basic Usage</h2>Task Launcher First-time setup <code>Run First Setup.bat</code> Validate before rendering <code>Run Validation Stress Test.bat</code> Preview one prompt, no render <code>Run One DRY TEST.bat</code> Base generation only <code>Run Nightrun BASE only.bat</code> Full run with scheduled upscale <code>Run Nightrun FULL upscale.bat</code> Re-upscale from metadata <code>Run 2x Upscale From Metadata.bat</code> Analyze prompt variety <code>Run Prompt Variety Analyzer.bat</code> Stop cleanly after current image <code>Stop After Current Image.bat</code><hr /><h2 id="metadata-upscaling">Metadata Upscaling</h2><pre><code>py upscale_from_metadata.py "D:\renders\0001_metadata.json" --config config.local.json --scale 2
</code></pre>Force base image as source:<pre><code>py upscale_from_metadata.py "D:\renders\0001_metadata.json" --config config.local.json --source base --scale 2
</code></pre>Use an existing upscale as source:<pre><code>py upscale_from_metadata.py "D:\renders\0001_metadata.json" --config config.local.json --source upscale --scale 2
</code></pre>Dry-run the upscale:<pre><code>py upscale_from_metadata.py "D:\renders\0001_metadata.json" --config config.local.json --scale 3 --dry-run
</code></pre><hr /><h2 id="first-run-setup">First-Run Setup</h2>The setup script creates a local config for your machine. It will:<ul><li>Ask for your Forge / Forge-Neo path and output folder</li><li>Detect your VRAM profile</li><li>Check checkpoint path and count local LoRAs</li><li>Check for Ultimate SD Upscale</li><li>Create <code>config.local.json</code> and desktop shortcuts</li></ul>The shared <code>config.json</code> stays portable. Your private machine paths stay in <code>config.local.json</code>. The project can be shared without leaking your local folder structure.<hr /><h2 id="recommended-for">Recommended For</h2>Nightrun is useful if you:<ul><li>Use Forge / Forge-Neo locally</li><li>Generate in large batches or overnight runs</li><li>Want editable prompt pools instead of hardcoded prompts</li><li>Use many LoRAs and want safer handling</li><li>Want metadata replay for good images</li><li>Want upscaling tied to generation metadata</li><li>Want less repetitive outputs over long runs</li><li>Want a workflow you can tune instead of replace</li></ul><hr /><h2 id="notes">Notes</h2>Nightrun does not include checkpoints, LoRA files, model files, Forge itself, or Ultimate SD Upscale. You need your own local Forge / Forge-Neo setup and your own models.Franchise and world module names are used as inspiration labels for prompt organization only. <h2 id="license-usage-note"> License / Usage note</h2><ul><li>Use this with models, LoRAs, and extensions you are allowed to use.</li><li>Franchise module names are inspiration labels for prompt organization. Nightrun does not include model files, LoRA files, checkpoints, or copyrighted assets.</li><li>Nightrun is a prompt generation and batch runner tool. It does not include, distribute, or download any AI models, checkpoints, LoRAs, extensions, or generated images. All content is produced locally using software and model files you have installed yourself.</li><li>You are solely responsible for the models and extensions you use, the content you generate, compliance with third-party license terms, and compliance with the laws of your jurisdiction.</li><li>NSFW generation is supported. It is your responsibility to ensure any explicit content you generate is legal in your jurisdiction. This tool must not be used to generate illegal content of any kind.</li><li>Nightrun is provided as-is, without warranty of any kind.</li></ul> <hr />Built for the dark hours between intention and discovery.

✦ NIGHTRUN ✦ Overnight Forge Randomizer (Low-High VRAM Optimized)

Nightrun 2.0 — ChangelogChanges from Nightrun 1.5Architecture — Monolith Split into ModulesIn 2.0, the engine was split from a single ~2055-line file into a modular system. The original <code>nightrun.py</code> is now a thin entrypoint that only imports and runs <code>main()</code>.The implementation is separated into purpose-driven modules:<ul><li><code>nightrun_runtime.py</code> — shared implementation (core logic)</li><li><code>prompt_engine.py</code> — pool loading, weighted selection, prompt composition</li><li><code>lora_resolver.py</code> — LoRA scanning, aliases, tag construction</li><li><code>forge_client.py</code> — Forge API, txt2img/img2img, GPU guards</li><li><code>runner.py</code> — execution modes, output folders, cleanup, cooldowns, summaries</li><li><code>validation.py</code> — pool validation and dry-run reporting</li><li><code>nightrun_engine.py</code> — compatibility shim for older imports</li></ul>New: Three-Engine Prompt RoutingPrompt generation now routes through three distinct builders, selected via weighted random choice (default 1/1/1):Classic — Standard pool sampling with optional world module support. Closest behavior to 1.5, cleaned and optimized.Combinator — Builds on Classic output by applying controlled transformations:<ul><li>Internal combinators modify existing elements (e.g. turning a katana into a burning variant)</li><li>External combinators inject additional motifs (e.g. surreal, cursed, steampunk layers)</li><li>At least one combinator may fire per prompt, with a cap of two.</li></ul>Franchise — Forces generation through a selected world module from <code>franchise_catalog.py</code>. All visual elements are sourced from that universe for full coherence. Falls back to Classic if no module is available.Routing can be controlled via <code>prompt_engines</code> in config, or disabled to revert to legacy behavior.Prompt Cleanup PipelineA post-processing cleanup stage was introduced via <code>prompt_cleanup</code> configuration. It handles:<ul><li>Duplicate comma-tag removal</li><li>Concept-family repetition control</li><li>Compact SD-style prompt length enforcement</li><li>Removal of redundant subject labels (e.g. "woman", "girl", "female") when already implied</li></ul>This ensures consistent prompt structure across all three engines.Prompt Pools — Full RebuildAll prompt pools were significantly reduced and rewritten for SD-style tag efficiency. The previous version contained large, verbose entries; 2.0 prioritizes compact, composable tags.Pool 1.5 lines 2.0 lines concept.txt 3,601 154 genre.txt 8,317 2,743 glamour_pose.txt 4,241 761 clothing.txt 1,801 721 character.txt 701 44 body_focus.txt 1,105 481 sensual_layer.txt 1,921 529 hair.txt 2,401 790Three pools were removed: <code>character_loras.txt</code>, <code>style_loras.txt</code>, <code>lora_sampler_speed.txt</code>Two new pools were added: <code>genre_core.txt</code> (69 lines), <code>genre_root.txt</code> (475 lines)A curated weighting system was introduced — <code>curated_top_count</code> and <code>curated_top_weight</code> — to boost high-quality early entries without removing full pool diversity.Franchise Catalog — ExpansionThe franchise system now uses full dictionary-based module definitions instead of compact tuples.The catalog expanded from ~24 to ~50 modules, including:<code>alien</code> <code>batman</code> <code>bloodborne</code> <code>cowboy_bebop</code> <code>doctor_who</code> <code>dungeons_and_dragons</code> <code>elder_scrolls</code> <code>evangelion</code> <code>final_fantasy_vii</code> <code>game_of_thrones</code> <code>halloween</code> <code>hellraiser</code> <code>lord_of_the_rings</code> <code>matrix</code> <code>midsommar</code> <code>nier_automata</code> <code>nightmare_on_elm_street</code> <code>persona</code> <code>pokemon</code> <code>scream</code> <code>shadowrun</code> <code>terminator</code> <code>the_conjuring</code> <code>the_ring</code> <code>the_witcher</code> <code>vampire_masquerade</code> <code>x_men</code> <code>zelda</code>New filtering controls allow fine-grained control without modifying the catalog:<ul><li><code>enabled_module_ids</code> / <code>disabled_module_ids</code></li><li><code>enabled_source_names</code> / <code>disabled_source_names</code></li><li><code>label_allow_patterns</code> / <code>label_block_patterns</code></li></ul>Selection Memory — RetuningSelection memory was expanded and softened to improve variation while preserving coherence.Key changes:<ul><li>Larger recent history windows across categories</li><li>Reduced repetition penalties for more natural variation</li><li>Added concept-level tracking (<code>recent_concepts</code>)</li><li>Dedicated repetition control for candle/light motifs</li><li>Curated weighting system introduced for high-quality biasing</li></ul>Config Changes<code>fixed_suffix</code> updated for sharper SD output consistency:<ul><li>Before: <code>clean coherent design, sharp focus, high detail, polished lighting</code></li><li>After: <code>crisp linework, sharp focus, high detail, clear material texture, clean color separation, high micro-contrast, dramatic lighting</code></li></ul>Other changes:<ul><li><code>base.steps_min</code> increased: 30 → 50</li><li>Output folders now use <code>nightrun-YYYYMMDD-HHMMSS</code> naming (legacy naming still supported)</li></ul>Setup PipelineA new VRAM profile was added — <code>tuned_6gb</code>:<ul><li>Resolution: 1344×768 base / 3072×1728 upscale</li><li>Steps: 65</li><li>LoRA cap: 7</li><li>VRAM threshold: ≤ 6144 MB</li></ul>Other changes:<ul><li><code>--reset-local</code> now preserves local tuning instead of overwriting it</li><li>Legacy setup shortcuts removed for cleaner initialization flow</li></ul>Sensual Layer — NSFW ExpansionThe <code>sensual_layer</code> pool was fully reworked to properly support explicit NSFW generation paths. The pool now includes dedicated nudity, erotic styling, body exposure, lingerie, fetish-adjacent aesthetic, and mature composition tags designed to integrate cleanly with the rebuilt SD-style prompt system.The expansion stays consistent with the 2.0 cleanup effort — shorter comma-tag formatting, improved SD readability, cleaner integration with prompt combinators, and stronger compatibility with franchise and genre routing.NSFW generation is enabled by default. For SFW-only runs, open <code>config.json</code>, locate the <code>sensual_layer</code> entry, and set <code>"chance": 0</code> (down from <code>0.24</code>).Final WordNightrun 2.0 is fully assembled into its modular form — restructured, refined, and expanded across prompt routing, pool design, and runtime behavior, while maintaining backward compatibility where needed. This marks a stable release point for the current architecture.Feedback, ideas, and bug reports are always welcome. The system is meant to evolve with use.

Files with SHA256: 0c8aa9858c59ab43f372f3273416cbf72342287fd064bb6901d2be0afbda4833