Production Nodes

comfyui-production-nodes

Four small ComfyUI nodes for the unglamorous part of generative production: running a batch from a brief, throwing out the bad frames, matching the grade, and proving later how each frame was made. Written against the V3 node API (comfy_api.latest), no extra dependencies, no models to download.

flowchart LR
    CSV[Prompt Batch<br/>from CSV] --> ENC[CLIPTextEncode] --> KS[KSampler] --> VAE[VAEDecode]
    CSV -- seed / cfg / steps / size --> KS
    VAE --> QG[Quality Gate]
    QG -- passed --> LUT[Apply .cube LUT] --> MAN[Save Images +<br/>Run Manifest]
    QG -- rejected --> PRE[Preview]
    CSV -- prompt / seed / tag --> MAN

| node | production problem it solves | |---|---| | Prompt Batch from CSV | Briefs arrive as spreadsheets. One row per queued run, typed outputs (prompt, negative, seed, cfg, steps, width, height, tag), errors that name the line and column. | | Quality Gate | Nobody should hand-review 400 frames to find the black, blown-out or mushy ones. Model-free metrics, configurable thresholds, a JSON report, and a graph that never breaks on an empty batch. | | Apply .cube LUT | Generated frames must sit next to graded footage. Applies the same .cube the colourist uses, trilinear on GPU, with intensity. | | Save Images + Run Manifest | "Which seed was the approved one?" A JSON sidecar per image with prompt, seed, params, a hash of the exact graph and a hash of the workflow structure. |

Synthetic chart (generated by scripts/make_assets.py): original, teal_orange, bleach_bypass.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/thejunglefruit/comfyui-production-nodes

Restart ComfyUI. The nodes appear under jungle_fruit/ in the node menu. Nothing to pip install: torch, numpy and Pillow already ship with ComfyUI. Requires a ComfyUI recent enough to expose comfy_api.latest (V3 nodes). Put your own LUTs in ComfyUI/models/luts/.

The repository also carries [tool.comfy] metadata in pyproject.toml for publishing to the Comfy Registry.

Nodes

Prompt Batch from CSV · JungleFruit_PromptBatchCSV

| inputs | outputs | |---|---| | csv_path (file in input/) or csv_text (pasted) | prompt, negative STRING | | index INT with increment control | seed, steps, width, height INT | | on_overflow: wrap / clamp / error | cfg FLOAT, tag STRING | | defaults for optional columns (advanced) | row_index, total_rows INT |

Required columns prompt, negative, seed; optional cfg, steps, width, height, tag. Queue the workflow N times and the index walks the sheet. Details

Quality Gate · JungleFruit_QualityGate

| inputs | outputs | |---|---| | images IMAGE | passed IMAGE | | min_sharpness (Laplacian variance) | rejected IMAGE | | min_brightness, max_brightness | report STRING (JSON) | | max_clipped_highlights, max_crushed_shadows | pass_count INT | | min_contrast, if_none_pass | |

Empty-batch policy. An empty branch is never emitted as a zero-length tensor; it is blocked silently, so nodes hanging off it skip that run without an error. If nothing passes, return_best (default) sends the best-scoring image through passed and flags "fallback_used": true in the report (pass_count stays 0); block skips the branch. Details

Apply .cube LUT · JungleFruit_ApplyCubeLUT

| inputs | outputs | |---|---| | image IMAGE (RGB or RGBA) | image IMAGE | | lut_name from models/luts and the pack's luts/ | | | intensity 0-1 | | | lut_path override (advanced, sandboxed) | |

Supports TITLE, LUT_3D_SIZE, DOMAIN_MIN/MAX, LUT_3D_INPUT_RANGE. Two example looks are included, computed from formulas so there is no third-party LUT licence involved. Details

Save Images + Run Manifest · JungleFruit_RunManifest

| inputs | outputs | |---|---| | images IMAGE, filename_prefix | manifest STRING (JSON) | | positive, negative, seed, tag | UI: saved images | | params (JSON or key=value lines) | | | include_graph, include_workflow | | | hidden: PROMPT, EXTRA_PNGINFO | |

Writes name_00001_.png + name_00001_.json. prompt_sha256 identifies the run; graph_sha256 ignores widget values and identifies the workflow. No host, user or path data is recorded. Details

Example

examples/production_batch_workflow.json (drag into ComfyUI) and examples/production_batch_api.json (for POST /prompt) wire all four nodes around a standard SDXL text-to-image graph. Copy examples/prompts.csv to ComfyUI/input/, pick your checkpoint, and queue as many runs as the sheet has rows.

Design

__init__.py              V3 entry point: ComfyExtension + comfy_entrypoint()
jf_nodes/core/           pure logic, no ComfyUI imports (CSV, metrics, LUT maths, manifest)
jf_nodes/nodes/          thin io.ComfyNode wrappers: schema, validation, caching, UI text
luts/                    example looks, generated by scripts/make_assets.py
scripts/                 asset generator and a ComfyUI load/smoke test
tests/                   pytest, CPU only

• Testable without ComfyUI or a GPU. All behaviour lives in jf_nodes/core; the node classes only translate between widgets and that logic.
• Caching that follows the files. The CSV and LUT nodes fingerprint file contents and modification time, so editing the sheet or the LUT re-runs the node; nothing else does.
• Paths are untrusted. File inputs must resolve inside ComfyUI's input/ or luts folders, which matters on shared or cloud ComfyUI servers.
• Readable failures. Every error names the node, the line or column, and what was expected.

Tests

pip install pytest
python -m pytest            # 55 tests, CPU only, no ComfyUI needed
python scripts/check_comfy_load.py /path/to/ComfyUI   # load + smoke-test inside a real install

The unit tests cover CSV parsing and errors, the metrics on synthetic sharp / blurred / burnt / black frames, LUT identity, a known inverting LUT, channel order, an independent numpy trilinear reference, domain handling, malformed .cube files, manifest hashing and atomic writes, and path sandboxing. The load script imports the pack the way ComfyUI does, builds every schema, validates the example API workflow with ComfyUI's own validator and runs each node once on synthetic data in a temporary folder.

Licence

GPL-3.0-or-later (see LICENSE).

Commercial licence. If you want to use it in a closed product without the GPL obligations, a commercial licence is available: [email protected]