SceneQueue

Generated with the included manga_demo preset.

SceneQueue turns large ComfyUI prompt batches into editable visual storyboards. Instead of maintaining a long JSON file, generate the whole sequence once and continue working directly from its gallery: edit any frame's prompt, compare the revision, re-queue it, or insert new frames before or after it without rebuilding the batch by hand.

Its optional AI assistant can use the surrounding prompts as narrative context to create up to five preceding, continuation, or transition frames. Every inserted prompt remains visible and editable, so the sequence grows under your control rather than becoming an opaque one-shot generation.

Important

Project status: SceneQueue is a portfolio and reference project. It is provided as-is under the MIT license, without active maintenance or support. Forks and independent adaptations are welcome.

The repository ships with general-purpose, safe-for-work examples and a workflow made only from standard ComfyUI nodes. Your prompts, generated images, selected models, and local workflow edits live under .data/ and are never committed.

See it in action

Mobile-first workflow Browse sessions, inspect generated frames, adjust generation settings, and manage prompt databases with touch-friendly controls.	AI-generated continuations Use AI Insert After to describe what happens next, generate multiple prompt segments, and queue the new sequence immediately.

Review before applying.
AI suggestions remain editable, and inserted text is highlighted before the updated prompt is accepted and re-queued.

Features

• Edit a generated batch directly from its visual gallery instead of maintaining a large prompt JSON file.
• Insert and generate frames before or after any existing image.
• Prompt revision history with visual diffs, editable AI suggestions, and re-queue.
• Context-aware AI generation of up to five preceding or continuation frames.
• Ordered prompt databases with insert, edit, history, and retry operations.
• Reusable placeholders for subject, appearance, wardrobe, pose, scene, and style.
• Sequential batch generation with configurable resolution, checkpoint, and LoRAs.
• Advanced sampler, scheduler, denoise, CFG, and optional high-resolution pass controls.
• Optional adult-content mode; general-audience negative tags are enabled by default.
• ComfyUI queue monitoring over WebSocket with HTTP fallback.
• Session gallery, thumbnails, failed-slot recovery, and optional upscaling.
• Optional OpenRouter assistant for prompt editing and transition-frame generation.
• Installable PWA frontend with no JavaScript build step.

Requirements

• Python 3.11 or newer.
• uv.
• A running ComfyUI instance reachable from this machine.

The default generation workflow uses these standard nodes: CheckpointLoaderSimple, CLIPTextEncode, EmptyLatentImage, KSampler, LatentUpscale, VAEDecode, and SaveImage. Selected LoRAs are injected with standard LoraLoader nodes at queue time.

Quickstart

git clone https://github.com/al4xdev/scenequeue.git
cd scenequeue
./install.sh
./start.sh

Open http://127.0.0.1:8889, then use Settings to configure:

• the ComfyUI URL;
• the local ComfyUI root directory;
• a checkpoint;
• the prompt node ID and input key if you replace the bundled workflow;
• generation resolution, checkpoint, and optional LoRAs.
• sampler and scheduler parameters;
• an optional high-resolution second pass;
• the general-audience/adult-content switch.

The Fast Illustration Sampling button applies a model-independent starting point: 12 base steps with the beta57 scheduler and a four-step high-resolution refinement. It preserves the checkpoint and LoRAs currently selected by the user.

The first run copies source defaults into .data/. Editing prompt databases or workflows through the application never changes the files committed in defaults/.

Included manga demo

Select the manga_demo database in each prompt category to queue an eight-frame story called The Last Delivery with its intended courier, wardrobe, station, and manga style. The sequence demonstrates:

• stable subject, wardrobe, prop, and location continuity;
• establishing, reaction, reveal, action, and closing shots;
• deliberate pacing across a complete visual micro-story;
• reusable placeholders, so the same sequence can feature a person, robot, fox, or any subject added by the user.

Docker

The latest multi-architecture image is published automatically to GitHub Container Registry for linux/amd64 and linux/arm64.

Warning

SceneQueue has no authentication layer. The examples below publish port 8889 to the local network so the mobile interface can reach it. Use only on a trusted network, or replace 8889:8889 with 127.0.0.1:8889:8889 for host-only access.

Docker Compose

docker compose up -d

Open http://localhost:8889. In Settings, use http://host.docker.internal:8188 as the ComfyUI URL when ComfyUI runs directly on the host. Change the port if your ComfyUI instance uses a different one.

The named scenequeue-data volume preserves configuration, prompt databases, workflows, API credentials, and gallery state across container upgrades.

To update:

docker compose pull
docker compose up -d

Docker CLI

docker pull ghcr.io/al4xdev/scenequeue:latest
docker volume create scenequeue-data
docker run -d \
  --name scenequeue \
  --restart unless-stopped \
  -p 8889:8889 \
  --add-host host.docker.internal:host-gateway \
  -v scenequeue-data:/app/.data \
  ghcr.io/al4xdev/scenequeue:latest

On Linux, host networking is also available:

docker run -d \
  --name scenequeue \
  --restart unless-stopped \
  --network host \
  -v scenequeue-data:/app/.data \
  ghcr.io/al4xdev/scenequeue:latest

With host networking, use the normal loopback ComfyUI URL, such as http://127.0.0.1:8188.

Build locally

docker build -t scenequeue .
docker run --rm -p 8889:8889 -v scenequeue-data:/app/.data scenequeue

Custom workflows

Export a workflow in ComfyUI's API format and replace:

• .data/workflows/workflow_api.json for generation;
• .data/workflows/upscale_api.json for upscaling.

Set target_node_id and target_input_key in Settings to the node/input that receives the positive prompt. Checkpoint and LoRA injection is discovered by node class, so users normally only need to choose another model in the interface.

The bundled upscale workflow expects RealESRGAN_x4plus.pth. Change its UpscaleModelLoader value if that model is not installed.

Optional OpenRouter assistant

Create .data/.env:

OPENROUTER_API_KEY=your-key
OPENROUTER_MODELS=deepseek/deepseek-chat,meta-llama/llama-3.3-70b-instruct

The API key remains server-side and is never returned by the settings endpoint. Without it, every non-AI editing and generation feature remains available. You can also configure or remove the key in Settings → OpenRouter Assistant; the password field is write-only and displays only whether a key is configured.

Data layout

.data/
├── config.json
├── databases/
├── gallery/
└── workflows/

Set SCENEQUEUE_DATA_DIR to store runtime data elsewhere. Set SCENEQUEUE_HOST and SCENEQUEUE_PORT to change the default 127.0.0.1:8889 listener.

This application has no authentication layer. Keep it bound to localhost or a trusted private network.

Development

uv sync --dev
uv run ruff check .
uv run mypy
uv run pytest

GitHub Actions runs the same lint and test checks without requiring a live ComfyUI server or OpenRouter account.

License

MIT