Branch a ComfyUI workflow on a natural-language judgment. Ask "which style does this prompt want?", "is this too graphic for kids?", or "does this name a real person?" and get a typed answer your graph can route on: a label, a probability, or a score, each with a confidence.

The nodes talk to three interchangeable System One decision models through the same question and answer shape. Clef can also look at images, which turns it into an automatic QA step for generated images.

| Backend | Where it runs | Cost | Setup | | --- | --- | --- | --- | | Laya | Locally, inside ComfyUI (MPS / CUDA / CPU) | Free | pip install laya | | Jev by TypeSafe | Hosted API | Per input token | TYPESAFE_API_KEY env var | | Clef / Clef-flash by Cloudflare | Workers AI API, sees images | Per input token ($0.09/M for flash) | CLOUDFLARE_ACCOUNT_ID + CLOUDFLARE_API_TOKEN env vars |

Flip one dropdown on the System One Backend node to move a whole graph between them.

Style router demo

New in 0.2.0: automatic image QA with Cloudflare Clef. Watch the 48-second intro.

Install

From the Comfy Registry: search for System One in ComfyUI Manager, or run

comfy node install comfyui-systemone

Or install from source:

cd ComfyUI/custom_nodes
git clone https://github.com/dante01yoon/ComfyUI-SystemOne
pip install -r ComfyUI-SystemOne/requirements.txt

For the hosted backends, export the keys in the shell that starts ComfyUI:

export TYPESAFE_API_KEY=...          # Jev
export CLOUDFLARE_ACCOUNT_ID=...     # Clef
export CLOUDFLARE_API_TOKEN=...      # a token with Workers AI permission

Laya downloads its weights from Hugging Face on first use and caches them. Loading takes about 6 seconds once per ComfyUI session; each answer after that takes 0.03 to 0.3 seconds on an Apple Silicon Mac.

Nodes

All nodes live under the SystemOne category.

| Node | Inputs | Outputs | | --- | --- | --- | | System One Backend | provider (laya / jev / clef-flash / clef), laya_checkpoint, device, jev_model | backend | | System One Choice | backend, state, instructions, options (one label: description per line), min_confidence, fallback | choice, confidence, confident, report | | System One Yes/No | backend, state, instructions, true_criteria, false_criteria, threshold | probability, verdict, report | | System One Score | backend, state, instructions, levels (one per line, low to high) | score, level, confidence, report | | System One Map | key, mapping (key => value lines, * => default) | value | | System One Image Check (Clef) | backend, images, instructions (a yes/no question about one image), reject_when (yes / no), threshold, context | passed (images that pass), passed_count, report | | System One Pick Best Image (Clef) | backend, images, instructions, context | best, index, confidence, report |

state is the text the model judges, usually your prompt. Text that parses as a JSON object or array is sent as structured state.
Choice falls back to fallback when confidence is below min_confidence, so an unsure model never picks a branch on a coin flip.
report is JSON with the question, the raw answer, the backend, and the latency, for logging or debugging.
Image Check asks its question about every image in the batch separately. If nothing passes, it blocks the nodes downstream, so nothing gets saved.
Pick Best Image compares up to 4 images per call and runs heats plus a final for bigger batches. index points into the batch it received.
Images go to Clef as downscaled JPEGs (long side 512 down to 256, to fit the Workers AI request budget). The passed and best outputs are your original full-resolution images.
After a run, each judgment node shows its answer and every option's probability as a text bar on the node itself.

Demos

Every demo below is a real recording of these workflows on ComfyUI with SDXL Turbo (4 steps, 512px). Each one ships as a drag-and-drop workflow in examples/. To run the image demos, put sd_xl_turbo_1.0_fp16.safetensors in ComfyUI/models/checkpoints.

1. Style router (Laya, local)

examples/style_router.json · video

A Choice node reads the prompt and picks photo, anime, watercolor, or render3d. A Map node turns the label into style keywords, which are appended to the prompt before sampling. With min_confidence at 0.5, an ambiguous prompt falls back to photo.

| Prompt | Laya answer | Routed to | | --- | --- | --- | | watercolor painting of a cat sleeping by the window | watercolor 0.98 (confidence 0.92) | watercolor | | cel-shaded anime girl with big eyes standing in the rain | anime 0.98 (confidence 0.90) | anime | | a knight riding a dragon over a city | anime 0.65 (confidence 0.26) | photo (fallback) |

| Clear style | Ambiguous, falls back | | --- | --- | | | |

2. Kid-safe gate (Laya, local)

examples/kid_safe_gate.json · video

Kid-safe gate demo

A Score node rates how graphic the requested image would be on four levels (none, mild, strong, extreme). A math node checks score >= 2, and an If/Else Switch swaps the prompt for a kid-safe replacement before anything is sampled, so a gory image is never generated.

On a 12-prompt test set, every safe prompt scored 1.39 or lower (a knight fighting a dragon, a pirate ship in a storm, a cartoon Halloween pumpkin) and every gore prompt scored 2.39 or higher, so the 2.0 cut separated all 12. The same question asked as a Yes/No got 11 of 12. In the recording, the gore prompt scored 2.31, which still clears the cut but by a smaller margin; treat the threshold as something to tune on your own prompts.

| Passes (score 1.48) | Swapped (score 2.31) | | --- | --- | | | |

3. Real-person guard: Laya vs Jev, side by side

examples/celebrity_guard.json · video

Celebrity guard demo

The same prompt goes to two Yes/No nodes, one per backend, asking whether it names a real celebrity or public figure. A yes swaps in a stand-in prompt. This demo stops before sampling on purpose, so it never renders a real person's likeness.

| Prompt | Laya | Jev | | --- | --- | --- | | Taylor Swift on stage, concert photo | 0.47, passes | 0.99, swapped | | Barack Obama eating ice cream on a bench | 0.36, passes | 0.99, swapped | | a businessman giving a speech on stage | 0.39, passes | 0.03, passes |

Laya's scores for real and generic people overlap (0.33 to 0.47 against 0.39), so no threshold separates them. Jev separates them by a wide margin. Use Jev for this kind of check.

4. One dropdown, two models

examples/model_switch.json · video

Model switch demo

Demo 1's graph, run twice per prompt: once with the backend set to Laya, then again after flipping it to Jev. These prompts imply a style through a reference instead of naming it.

| Prompt | Laya | Jev | | --- | --- | --- | | a girl on a grassy hill under huge summer clouds, in the style of Studio Ghibli | anime, confidence 0.07, falls back to photo | anime, confidence 1.00 | | a lighthouse at dusk, loose wet-on-wet brushwork | watercolor, confidence 0.19, falls back to photo | watercolor, confidence 1.00 |

| Laya | Jev | | --- | --- | | | | | | |

On 8 such prompts (Ghibli, Pixar, Makoto Shinkai, Unreal Engine 5, Canon 5D, National Geographic, wet-on-wet), Jev answered all 8 correctly with confidence 0.99 or higher. Laya's top answer was right on 7, but 4 of those had confidence below 0.5, so with demo 1's gate Laya routes 3 of 8 to the intended style.

5. Automated image QA (Clef-flash)

examples/image_qa.json · video

Image QA demo

SDXL Turbo renders 4 candidates for "exactly three red apples on a white table". Image Check asks Clef-flash about each one: "Does the image show exactly three apples, no more and no fewer?" Images that fail are dropped. Pick Best Image then chooses among the survivors, and only that one is saved. Nobody has to look through the batch.

The four candidates

| Candidate | Apples | Clef-flash "exactly three?" | Result | | --- | --- | --- | --- | | 1 | 3 | 0.96 | pass | | 2 | 3 | 0.97 | pass | | 3 | 3 | 0.97 | pass, picked as best (0.79) | | 4 | 4 | 0.03 | rejected |

A second seed in the recording also caught the one bad image (five apples) and kept three; the pick among three near-identical survivors came back with low confidence (0.16), which is a cue that any of them would do.

Clef-flash really reads the image rather than the prompt. Asked "does this image match a watercolor cat?" about a photo of a knight, it said 0.004; the same question with no image attached said 0.90. Given four different images and four prompts, it picked the right image for each, with confidence 0.92 to 0.96, in 0.3 to 0.9 seconds per call.

Seed 7 run

Choosing a backend

Measured on 2026-10-02 with laya 0.3.24 (convaiinnovations/laya, MPS) and jev-1.13.0:

| Task | Laya | Jev | | --- | --- | --- | | English prompt that names its style | Correct, confidence 0.5 to 0.99 | Correct, confidence 1.00 | | Style implied by a reference (Ghibli, Pixar) | Mostly right, low confidence | Correct, confidence 0.99+ | | Korean prompt ("수채화", "셀 셰이딩 애니메이션") | Wrong | Correct | | Names a real celebrity or public figure | Cannot separate | 0.99 vs 0.03 | | Graphic or gory content (Score) | Separated 12 of 12 at 2.0 | Not measured | | Latency per answer | 0.03 to 0.3 s, local | 0.15 to 0.23 s, network | | Looks at images | No | No |

Start with Laya for explicit, English, local judgments. Switch to Jev when the answer needs world knowledge (people, artists, studios) or another language. Use Clef when the question is about the generated image itself; it answered the image questions above in 0.3 to 0.9 s.

Re-recording the demos

The recordings come from two scripts, so you can reproduce or change them.

scripts/run_demo.py queues an API-format workflow once per prompt and prints each node's answer. It needs only Python and a running ComfyUI.
scripts/record_demo.mjs opens ComfyUI in Playwright, lays out the nodes, types each prompt, clicks Run, and records the screen. An argument like @2.provider=jev changes a widget between runs. scripts/make_media.sh turns the recording into an MP4 and a README GIF.

python scripts/run_demo.py examples/style_router.api.json --prompt-node 3 \
  --prompts "watercolor painting of a cat" "a knight riding a dragon over a city"

npm install
node scripts/record_demo.mjs examples/model_switch.api.json 3 \
  "a lighthouse at dusk, loose wet-on-wet brushwork" "@2.provider=jev" \
  "a lighthouse at dusk, loose wet-on-wet brushwork"
START=2 scripts/make_media.sh model_switch

Both scripts expect ComfyUI on http://127.0.0.1:8199. Override it with --base or COMFY_URL.

Limits

Laya and Jev judge text only; give image questions to a Clef backend.
Workers AI estimates a request's size from its raw body, so Clef calls carry at most 4 images and downscale them to fit. A 512px PNG batch of four is rejected before the model sees it; the nodes handle this for you.
Laya reads at most 512 tokens of state, and its option labels share a 192-token budget.
Laya's checkpoint warns that some of its temperatures are outside the expected range, so treat its confidence as uncalibrated.
SDXL Turbo, used in the demos, is under a non-commercial research license. The nodes themselves are MIT.

Tests

python -m pytest tests -q

The tests use a fake backend, so they need neither model nor network.