Extensions/ComfyUI-Goofer
ComfyUI Extension

ComfyUI-Goofer

Random IMDb movie goof → AI video prompt → LTX-2 video clips → procedural interstitial → MusicGen film score → final stitched video.

By jbrick2070·Created 5 months ago·Updated 5 months ago· 2
jbrick2070/ComfyUI-Goofer
Nodes—
On cloudLocal install
Stars2
Updated5 months ago
Readme

ComfyUI-Goofer v1.1

IMDb Goof-to-Video Pipeline

Random IMDb movie goof → AI video prompt → LTX-2 video clips → procedural interstitial → MusicGen film score → final stitched video.

Fully automated. Zero API keys. Drop into custom_nodes/ and queue.

Download

Download Goofer v1.1

Click here to download the full package (v1.1) — includes workflow JSON + this guide.

What It Does

Goofer picks a random film from a pool of 424 curated titles, fetches its IMDb continuity errors and goofs, sanitizes all copyrighted names and brands, generates a cinematic video prompt for each goof, renders an LTX-2 video clip per goof, creates a procedural data-viz interstitial with an AI-generated film score, then stitches everything into a single upscaled output video.

Every run is a different movie. Every queue is a new short film.

New to ComfyUI? Start Here

ComfyUI is a free, node-based interface for running AI image and video models locally on your GPU.

Already have ComfyUI installed? Skip to Step 2.

Step 1 — Install ComfyUI

Use the official desktop installer — it handles Python, Git, and dependencies automatically:

https://www.comfy.org/download

Advanced users can also install manually from https://github.com/comfyanonymous/ComfyUI

Step 2 — Install Required Models

☕ Grab a coffee — the main video model is ~9.5 GB.

| Model | Download | Size | Save To | |-------|----------|------|---------| | LTX-Video v0.9.5 (required) | HuggingFace | ~9.5 GB | ComfyUI/models/checkpoints/ | | LTX-Video 13B (optional, higher quality) | HuggingFace | ~26 GB | ComfyUI/models/checkpoints/ | | Qwen2.5-3B-Instruct (default AI) | Auto-downloaded on first run | ~6 GB | HuggingFace cache | | Qwen2.5-7B-Instruct (optional, higher quality) | Auto-downloaded on first run | ~14 GB | HuggingFace cache | | MusicGen Large | Auto-downloaded on first run | ~3.3 GB | HuggingFace cache |

Qwen2.5 and MusicGen download automatically the first time you run the workflow — no manual steps needed.

Step 3 — Install ComfyUI-Goofer

Option A — ComfyUI Manager (recommended):

  1. Open ComfyUI Manager → Install Custom Nodes
  2. Search "Goofer" → Install → Restart

Option B — Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/jbrick2070/ComfyUI-Goofer

Then install Python dependencies:

pip install cinemagoer requests transformers torch

Step 4 — Load the Workflow

Goofer ships with two workflow presets:

| Workflow | Target | Hardware | File | |----------|--------|----------|------| | Basic | Local Windows | RTX 5080 / 4090 | example_workflows/goofer_basic.json | | Blackwell RunPod | RunPod Linux | RTX PRO 4500 (Blackwell) | example_workflows/goofer_blackwell_runpod.json |

Both workflows use the same node pipeline and produce identical output. The Blackwell variant is optimized for RunPod Linux with 175-frame clips (5 seconds at 35 FPS), TensorRT-accelerated NVIDIA upscaling via tensor cores, and high-performance text encoding.

  1. Open ComfyUI at http://127.0.0.1:8000
  2. Click Load and select the workflow for your environment
  3. If any nodes appear red: Manager → Install Missing Custom Nodes → Restart
  4. Hit Queue — Goofer picks a random movie and starts generating

Which GPU do I need?

| Setup | GPU | VRAM | |-------|-----|------| | Recommended (local) | RTX 5080 / 4090 | 16 GB+ | | Recommended (cloud) | RTX PRO 4500 Blackwell (RunPod) | 24 GB+ | | Minimum | RTX 4070 / 3060 | 8 GB |

Check your VRAM: Windows: Task Manager (Ctrl+Shift+Esc) → Performance → GPU → Dedicated GPU Memory. Linux: nvidia-smi

Note on 8GB VRAM: v1.1 increased the default clip generation resolution to 1024x576 for true 16:9 widescreen output. If you encounter CUDA Out of Memory errors during rendering, drop the width and height widgets on GooferBatchVideo and GooferProceduralClip back down to 768 and 512.

The Node Pipeline

GooferInit
    │
    ▼
GooferGoofFetch ──────────────────────────────────┐
    │  movie_data  goofs_data                      │
    ▼                                              │
GooferSanitizer                                   │
    │  sanitized_goofs                             │
    ▼                                              │
GooferPromptGen                                   │
    │  prompts (×5)                                │
    ▼                                              │
GooferBatchVideo                                  │
    │  video clips (×5)                            │
    ▼                                              ▼
 [final stitch] ◄── GooferProceduralClip ◄── GooferBackgroundMusic

Node Reference

| Node | What It Does | |------|-------------| | GooferInit | Sets global config: output path, upscale mode, seed | | GooferGoofFetch | Picks a random film from 424 titles, fetches goofs from IMDb via Cinemagoer + direct HTTP fallback. Seed-diversified so every run draws a fresh subset from the full goof pool | | GooferSanitizer (Copyright Cleaner) | Strips copyrighted names, brands, and franchises. Now includes the robust Banana Filter which humorously intercepts and replaces all firearms, heavy weaponry, and violent verbs with bananas before reaching any AI model | | GooferPromptGen | Converts sanitized goofs into cinematic LTX-2 prompts via Qwen2.5 (3B default, 7B optional). Forces the exact mistake to be the central focal point of the scene and dynamically injects cinematic tracking/lighting highlight styles to explicitly point out the error. Lazy-loads and unloads the model to free VRAM before video generation | | GooferBatchVideo | Feeds each prompt to LTX-Video and renders one video clip per goof. v1.1 defaults to native 16:9 output at 1024x576, alongside configurable frame counts and guidance strengths | | GooferBackgroundMusic | Generates a film score with Meta MusicGen 3. Prompt is derived from actual goof keywords (explosions → percussive bass, chases → driving rhythm, basketball → upbeat brass, etc.) plus Flan-T5-small genre/mood inference from the film's plot. Falls back to additive-synthesis chord progression if MusicGen is unavailable | | GooferProceduralClip | Renders a data-viz interstitial showing goof metadata as animated neon graphics. Duration syncs to the MusicGen output length via the music_duration input | | GooferVideoConcat (Stitch + Upscale) | Concatenates up to 6 VIDEO clips with optional crossfade dissolve transitions. 4-tier upscaler: nvvfx RTX VSR (Windows), TensorRT-accelerated Real-ESRGAN (Linux/RunPod — NVIDIA tensor cores), standard Real-ESRGAN (fallback), bicubic (last resort). Default target: 1080p | | GooferAudioEnhance | Optional EQ / loudness normalization pass on the final audio mix |

Movie Pool

Goofer ships with 424 curated titles spanning 1939–2024:

Action · Sci-Fi · Horror · Comedy · Drama · Western · War · Disaster · Spy/Espionage · Animation · Romance · Musical · Sports · Cult Classics

Each entry is verified against IMDb with a title-sanity check — if an ID returns the wrong film, Goofer skips it and picks another rather than generating content for the wrong movie.

To add your own films, edit _RANDOM_MOVIE_POOL in goofer_goof_fetch.py:

("Your Movie Title", YEAR, "IMDB_ID_WITHOUT_TT"),
# e.g.
("The Thing",        1982, "0084787"),

Copyright Safety

Goofer is designed to keep copyrighted names out of AI-generated video prompts. The Copyright Cleaner (GooferSanitizer) node runs several passes before any text reaches LTX-Video or MusicGen:

  • Franchise names replaced with generic descriptions ("James Bond" → "a spy thriller")
  • Movie title as character name → "the character"
  • Movie title as title reference → "the film"
  • Actor and character names from the IMDb cast list stripped automatically
  • Studio and production company names stripped
  • Brand names including model names ("Lamborghini Gallardo") → "a well-known brand"
  • Standard First-Last name patterns detected and replaced

Troubleshooting

<details> <summary><strong>Nodes load with wrong widget values / "Failed to convert FLOAT"</strong></summary>

Cause: Stale .pyc bytecode in __pycache__ loaded before the updated .py files.

Fix: Delete the cache folder and restart ComfyUI:

Remove-Item "ComfyUI/custom_nodes/ComfyUI-Goofer/__pycache__" -Recurse -Force
</details> <details> <summary><strong>"No goofs found" / placeholder goofs appearing</strong></summary>

Cause: Cinemagoer's HTML parser is broken against IMDb's current page structure (known upstream issue). Goofer includes a direct HTTP fallback using __NEXT_DATA__ JSON extraction, but some films legitimately have zero goofs listed.

Fix: The fallback is automatic. If a specific film keeps returning empty, verify its IMDb ID by checking https://www.imdb.com/title/ttXXXXXXX/goofs/ in your browser.

</details> <details> <summary><strong>MusicGen prompt shows "cinematic atmospheric" every run</strong></summary>

Cause: No plot text from Cinemagoer (common on first fetch for some titles), so Flan-T5 skips inference and the genre keyword dict has no match.

What to expect: This is normal for some titles on first run. Goof keyword matching (explosions, chases, fights, sports, etc.) still enriches the prompt regardless of genre availability.

</details> <details> <summary><strong>MusicGen takes a long time on first run</strong></summary>

Cause: facebook/musicgen-large (~3.3 GB) downloads on first use and loads into VRAM alongside LTX-Video.

Fix: One-time only. After first run both models stay cached locally. If VRAM is tight, switch to facebook/musicgen-small (~300 MB) in the GooferBackgroundMusic node widget.

</details> <details> <summary><strong>"Title mismatch" warning in logs</strong></summary>

Cause: The IMDb ID in _RANDOM_MOVIE_POOL points to the wrong film (reused ID, sequel confusion, or regional title difference).

Fix: Look up the correct ID at https://www.imdb.com/find?q=TITLE+YEAR and update the 8-digit number in goofer_goof_fetch.py.

</details>

Setting Up Continuous Output with OBS

Run Goofer as a live generative broadcast — each output video auto-loads into OBS as it finishes.

Prerequisites

Setup

  1. Install OBS and the Media Playlist Source plugin
  2. In OBS: Tools → Scripts → Python Settings → point to your Python 3.11 path
  3. Load the directory_sorter_for_obs script and point it to ComfyUI/output/
  4. Add a Media Playlist Source scene item pointed to the same folder
  5. OBS automatically picks up each new Goofer output as it renders

Pro tip: If your main GPU is maxed out on inference, set OBS to encode via your integrated GPU (QSV AV1 or HEVC). This keeps the stream smooth while the RTX handles LTX-Video and MusicGen.

Content Policy � No NSFW / Explicit Content

ComfyUI-Goofer includes a mandatory two-stage explicit content filter built into the Copyright Cleaner (GooferSanitizer) node. It runs on every goof before any text reaches LTX-Video or MusicGen.

Stage 1 � Keyword & Pattern Block (always on) Any goof containing explicit sexual terms, nudity references, graphic gore, or child-safety violations is silently dropped and logged. It never reaches the prompt generator.

Stage 2 � AI Judge (Flan-T5, default on) Borderline goofs containing words like "bare", "intimate", "strip", or "shower" are passed to Flan-T5-small with the instruction: "Is this description appropriate for all audiences and free from sexual or explicit content?" If the answer is "no", the goof is dropped.

This filter cannot be disabled from the workflow UI. The only override is editing the source directly, which you should not do unless you are using Goofer with fully public-domain films in a verified adult-only context compliant with all applicable laws.

Goofer is intended for general audiences. Do not attempt to circumvent this filter.

Requirements

cinemagoer
transformers>=4.37.0
accelerate>=0.26.0
torch>=2.0.0
numpy
Pillow>=10.0.0
opencv-python>=4.8.0

For NVIDIA upscaling (recommended):

realesrgan>=0.3.0
basicsr>=1.4.2

For RunPod / Linux Blackwell (TensorRT tensor core acceleration):

torch-tensorrt>=2.2.0

Optional (for higher-quality audio resampling):

librosa

Built by Jeffrey A. Brick — Los Angeles, 2026