ComfyUI VLM Nodes

Production-oriented vision-language, structured prompting, audio, and utility nodes for ComfyUI. Version 3.4 supports ComfyUI's selected NVIDIA CUDA, AMD ROCm, Apple Metal, Intel XPU, and CPU device without replacing its PyTorch build. It removes startup installers and global accelerator cache flushes, adds real image/video batches and live token streaming, and uses ComfyUI model residency and offloading.

Modern model coverage

The Modern VLM node provides one stable interface with a deliberately small, 12-choice production picker:

• Qwen 3.5 0.8B and 4B
• Qwen 3 VL 2B, 4B, and 8B Instruct
• SmolVLM2 500M and 2.2B Video
• Liquid LFM2.5-VL 450M
• InternVL 3.5 1B
• Granite Vision 4.1 4B
• Gemma 3 4B IT
• a compatible custom Hugging Face image-to-text repository

The separate [Legacy] Modern VLM Compatibility node contains redundant, superseded, experimental, and very large tiers:

• Qwen 3.5 2B, 9B, 27B, and 35B-A3B
• Qwen 3.6 27B
• Qwen 3 VL 30B-A3B Instruct
• Qwen 2.5 VL 3B and 7B for existing workflows
• Gemma 3 12B and 27B IT
• SmolVLM2 256M Video
• Liquid LFM2.5-VL 1.6B
• InternVL 3.5 2B
• Granite Vision 3.3 2B

Previously saved ModernVLM workflows remain valid even when their selected model moved to Legacy. The server accepts every known catalog value for backward compatibility; only the visible new-workflow picker is curated. Dedicated Molmo, PaLI-Gemma, Qwen2-VL, MiniCPM-V, Kosmos-2, MC-LLaVA, UForm, and script-style MoonDream nodes are also collected under VLM Nodes/Legacy/Model Loaders. Maintained creator-facing Florence-2, Moondream2, JoyTag, llama.cpp/GGUF, detection, segmentation, tracking, API, and video-intelligence nodes stay in their functional categories.

Sixteen curated sub-4B/low-VRAM choices are marked internally as the small-and-fast tier. The default is Qwen 3 VL 2B: it is much quicker to load than larger checkpoints while retaining broad image and video understanding. The catalog intentionally uses official model repositories and maintained Transformers interfaces rather than unverified community quantizations. Curated models use native Transformers implementations; remote repository code is enabled only when the explicit custom-model option requires it. Florence-2 uses the Transformers-native converted checkpoints instead of Microsoft’s legacy repository code.

Live text output

Modern VLM streams decoded text through ComfyUI's native progress_text WebSocket channel by default. A connected ViewText node updates while tokens arrive, shows the final response after execution, and restores the last result when ComfyUI rehydrates workflow output history. Disable stream_output for API-only or headless runs that do not need incremental UI updates. Streaming is best-effort and never changes the final STRING output or makes inference fail.

Text workflow toolkit

The original SimpleText, JsonToText, and ViewText node IDs and their first STRING outputs remain stable for saved workflows. They now live in organized VLM Nodes/Text subcategories and expose descriptive names, search aliases, tooltips, appended metrics, and strict error messages:

The JSON utilities never evaluate code, follow references, access files, or make network requests. Template fields are direct names rather than Python attribute/index expressions. approx_tokens is deliberately labeled as a rough UTF-8 budget estimate; use the target model tokenizer when exact billing or context accounting matters.

Specialized nodes remain available where a generic chat node would discard useful model capabilities:

• Moondream 3.1 9B-A2B: official 2B-active Photon runtime with query, caption, and high-throughput image/video detection and pointing.
• Moondream 3 Preview segment: native SVG segmentation through the same isolated Photon loader. The SVG is preserved and also converted into antialiased MASK, black/white previews, foreground cutouts, overlays, polygons, canonical VLM_DETECTIONS, and core bounding boxes. Detection/pointing submit frames concurrently so Photon can dynamically batch them; every run reports measured worker FPS, end-to-end FPS, and real-time factor.
• Florence-2: captioning, OCR, detection, region captioning, and referring expression segmentation, with structured JSON, mask, and overlay outputs.
• PaLI-Gemma: caption/VQA plus the official 16-token VQ-VAE segmentation decoder; segmentation tokens are no longer misinterpreted as polygon points.
• Moondream2: pinned query API with explicit decoding controls. The official checkpoint is loaded through its native safetensors state dict, avoiding the silent empty-output regression in Transformers 5 while retaining ComfyUI managed loading and unloading.
• Qwen2-VL: image batches and real video-frame batches.
• Legacy Molmo, Kosmos-2, UForm, MCLLaVA, and MiniCPM-V 2.6 GGUF, plus maintained JoyTag.
• llama.cpp LLaVA/GGUF, structured prompt suggestions, OpenAI-compatible prompting, and AudioLDM2.

Structured detection, segmentation, and tracking

The vision nodes use stable, typed sockets instead of passing model-specific lists between nodes:

All spatial coordinates are source-image pixels. Bounding boxes are [x1, y1, x2, y2] with an exclusive right/bottom edge; polygons contain at least three points and quads exactly four. JSON roots contain schema, version, media dimensions/frame count/FPS, and their ordered records. Dense mask tensors remain in-process and are deliberately omitted from JSON so API results do not unexpectedly grow by hundreds of megabytes.

The utility layer converts without model-specific glue:

• VLMStructuredSpatialParser strictly parses pixel, normalized 0–1, or normalized 0–1000 JSON from any VLM into VLM_DETECTIONS and VLM_POINTS. VLMSpatialPromptBuilder creates the matching constrained prompt.
• VLMDetectionsToBoundingBoxes, VLMDetectionsToPoints, and VLMDetectionsToMasks emit Comfy core boxes, center points, combined and individual binary masks, inverse masks, ready-to-preview black-and-white images, and stable-color instance maps. Polygon/quad masks are rasterized when present, otherwise the bounding box is used. Existing output indexes remain stable; the creator-facing mask images and instance map are appended.
• VLMFilterDetections, VLMSelectDetection, VLMCropDetections, and VLMRenderDetections provide label/score/area/frame selection, padded crops, and deterministic overlays.
• VLMMaskProcessor accepts any Comfy MASK, including SAM2/SAM3 masks, and returns a feathered matte, strict binary mask, inverse mask, and black-and-white image. Its grow/shrink and Gaussian feathering run in Torch without OpenCV or SciPy.
• VLMMaskComposite applies still-image or video mask batches to a source and returns the replacement composite, isolated foreground, original background-only plate, and black-and-white mask image. A single mask or background broadcasts safely across a video batch.
• VLMDetectionsFromJSON and VLMDetectionsToJSON are the explicit API and persistence boundary for the versioned detection schema.

Universal VLM performance utilities

The performance nodes sit before any local or hosted VLM, so their savings do not depend on CUDA, ROCm, MPS, XPU, CPU, Transformers, llama.cpp, or Photon:

• VLM Performance Profile emits coherent max_frames, pixel budget, longest-edge, batch-size, and unload_after values. Live / robotics, Fast video, Balanced, High detail, and Low VRAM handoff are explicit starting points rather than hidden global flags.
• VLM Adaptive Frame Sampler is the existing track-aware temporal gate. It combines uniform coverage, scene changes, motion, and optional track changes while preserving source frame indices and timestamps.
• VLM Image Pixel Budget downsizes the selected analysis copy once, preserves aspect ratio, never upscales, and can align dimensions to 14/28-pixel VLM patches or 32-pixel detector backbones. Fast area and antialiased bicubic modes are available.

The recommended order is Video Slice → VLM Adaptive Frame Sampler → VLM Image Pixel Budget → any VLM. A model's own official processor still performs its required normalization/crop; the pixel-budget node simply prevents every downstream model from repeatedly receiving unnecessary source pixels. Local torch models remain registered with ComfyUI's smart model manager, while external allocators reserve space before loading and close only the handle they own.

On the real vlm_api_people_birds.mp4 input in this repository's D-drive test environment, the utilities selected 10 of 60 1280×720 frames and resized them to 938×518 in about 0.44 seconds on a cold WSL run. That reduced the frame×pixel analysis workload by 11.38× before model inference. This is an input-work reduction measurement, not a claim that every model runs 11.38× faster; token generation and model-specific vision encoders still determine end-to-end speed.

Adaptive video intelligence

The video-intelligence layer keeps generative VLM inference out of the per-frame loop:

• VLMAdaptiveFrameSampler combines scene-change, motion, track-change, and uniform-coverage signals. It always preserves the real source frame index and timestamp, enforces a frame budget, and returns selection/diagnostic JSON. Uniform coverage, motion, scene, and track-priority modes remain available for deterministic experiments.
• VLMVideoTemporalReasoner is the one-node path. It adaptively samples the input, downsizes only the VLM analysis copy (448-pixel longest side by default), runs a recommended video-capable model, parses the result into validated VLM_EVENTS, and returns summary, events, selection, sampled previews, raw response, diagnostics, event JSON, and selection JSON.
• VLMVideoReasoningPrompt and VLMEventsFromVideoJSON expose the same strict timestamp/evidence contract for custom local or hosted VLM workflows.
• VLMTrackAwareCrops chooses representative observations for each durable track, adds configurable context, and letterboxes crops to one batch size. This lets a VLM label identities without rereading every full frame.
• VLMBuildSceneState converts tracks plus optional events into a compact persistent world-state summary with first/last observation, current box, confidence, state, and pixel velocity.

Small VLMs commonly return evidence as positions in the supplied image batch even when asked for source indices. The parser accepts that form only when every value is an unambiguous valid supplied-image position, maps it back to the immutable source selection, and records the normalization mode. Arbitrary or unsupplied evidence frames, out-of-range timestamps, invalid confidence, duplicate evidence, malformed JSON, and non-finite values fail validation.

On the repository's real-data smoke test (RTX 3090, Qwen3-VL 2B, 157-frame 896x448 H.264 clip), hybrid sampling selected 12 frames in 0.30 seconds, reduced temporal inputs by 92.36%, reduced analysis pixels by 75%, used 4.24 GiB peak allocated VRAM in the standalone runner, and produced a valid timestamped result in 35.17 seconds. The equivalent live ComfyUI /prompt graph completed in 37.45 seconds. These are one-machine measurements, not portable performance guarantees.

Open-vocabulary image and video detection

VLMOpenVocabularyDetection exposes one interface for:

• Grounding DINO Tiny and Base
• OWLv2 Base Ensemble
• OmDet Turbo Swin Tiny

It accepts a still image or an IMAGE batch of video frames and processes the batch frame by frame. Outputs, in socket order, are detections, json, preview, box_mask, and Comfy core bounding_boxes. Connect the FPS output of GetVideoComponents when the input is video so every timestamp is correct. For tracking-by-detection, run detection over the complete bounded batch and connect it to VLMTrackDetections.

VLMTrackDetections uses a ByteTrack-style two-stage high/low-confidence association, motion prediction, label-aware matching, and time-based expiry. IDs are durable within the supplied sequence and survive short missed detections when emit_predictions is enabled. Independent Comfy queue runs or independently sliced chunks are separate tracking sessions; they do not silently reuse IDs.

SAM2.1 and Comfy core SAM3.1

VLMSAM2VideoSegmentation propagates first-frame detections, one core BOUNDING_BOX, or seed masks through an IMAGE batch using SAM2.1 Hiera Tiny, Small, Base+, or Large. It returns VLM_TRACKS, report JSON, per-frame union masks, frame-major individual object masks, and an overlay batch. The object IDs assigned at the seed frame remain stable for that video session.

VLMSAM3TrackAdapter is intentionally an adapter, not a second SAM3 loader. It validates ComfyUI core SAM3_TRACK_DATA, preserves the core bit-packed mask payload unchanged, and exposes lightweight VLM_TRACKS metadata with mask references. Connect its passthrough output to core SAM3_TrackPreview or SAM3_TrackToMask, and connect tracks to VLMTrackReport. This avoids duplicating dense masks in memory or JSON.

SAM3 weights use Meta's SAM License. The upstream facebook/sam3 repository requires accepting access terms and sharing the requested account information; the ComfyUI checkpoint is also marked sam-license. Review and accept the license before downloading. The example names ComfyUI's sam3.1_multiplex_fp16.safetensors; if it is unavailable, use the SAM2.1 workflow rather than substituting an unrelated checkpoint.

Florence-2 task coverage

Florence2 exposes all 15 supported task contracts:

Every task returns text, structured_json, mask, and visualization. Tasks that do not produce a spatial result return an empty mask and the source image visualization. Region tasks reject ambiguous multi-box input; use VLMSelectDetection to isolate the record, then supply exactly one core BOUNDING_BOX with the same pixel coordinates.

Video memory strategy

• Trim long media with core Video Slice, then use GetVideoComponents. Downscale the complete frame batch before detection or segmentation and keep every frame at identical dimensions.
• Grounding detection supports configurable micro-batches; keep batch_size=1 for minimum VRAM or increase it when memory allows. It returns both nested per-frame core BOUNDING_BOX values and flat metadata-rich BOUNDING_BOXES.
• SAM2.1 stores source video frames on CPU, keeps its inference state on CPU by default, and limits the vision-feature cache to one frame. Union masks and previews return on CPU. Full per-object mask volumes are opt-in with mask_output=union_and_objects; disable render_preview to avoid another full-resolution overlay copy on long clips.
• Start with Grounding DINO Tiny plus SAM2.1 Hiera Tiny. Increase detector or segmenter size only after the pipeline is correct. unload_after=false caches one model per node instance; use true when another large model must run immediately afterward.
• A Video Slice is an independent propagation session. For very long media, use bounded slices, reseed each slice, and keep the overlap/output mapping in the caller. The pack does not pretend IDs are globally stable across separate queues.
• The SAM3 adapter never unpacks the complete mask volume for its report. Use core SAM3_TrackToMask only when a dense selected mask is actually needed.

API-format examples are in examples/vision:

• grounding_dino_image_api.json
• moondream3_preview_svg_segment_api.json
• moondream31_video_detect_api.json
• sam2_video_tracking_api.json
• sam3_core_adapter_blueprint_api.json
• video_temporal_reasoning_api.json
• vlm_performance_preflight_api.json

The dependency-free text-toolkit example is examples/text_toolkit_api.json. Robotics policy, safety, and sidecar examples are in examples/robotics, including a complete universal HTTP policy graph.

Upload the named media to ComfyUI's input directory, adjust the filenames and labels, then submit the JSON object as the prompt value to /prompt. These are API graphs, not frontend workflow-export JSON.

Node reference

All 89 registered nodes, grouped by their menu category. The Node ID is the class_type written into workflow and API JSON — search for that string when you need to find a node you saw on a canvas.

Modern VLM

The main entry point for current vision-language models.

Moondream 3

Moondream 3 / 3.1 in an isolated Photon runtime. Load once, then reuse the MOONDREAM31_MODEL output across the task nodes.

Florence-2

Vision: detection, segmentation, tracking

Open-vocabulary detection and video segmentation. These emit the structured VLM_DETECTIONS / VLM_POINTS / VLM_TRACKS types rather than loose strings.

Vision: spatial reasoning

Vision: detection utilities

Converters and filters between structured detections and ordinary Comfy types.

Vision: mask tools

Video intelligence

Adaptive frame selection and temporal reasoning for long videos.

LLM (local GGUF)

llama.cpp text models. LLM Loader (GGUF) produces the CUSTOM model handle the samplers consume; the Managed Cache variants own their own handle and can release it after each run.

LLaVA (local GGUF multimodal)

Vision models through llama.cpp. These need both a GGUF and its vision projector (mmproj).

Hosted APIs

Robotics / VLA policies

These nodes build and inspect policy observations/actions. They never send commands to robot hardware. Heavy policy runtimes stay in isolated LeRobot, openpi, GR00T, OpenVLA/OFT, or JAX environments.

Text toolkit

Dependency-free string handling, so a VLM response can be shaped without an extra node pack.

Performance and diagnostics

Run VLM Runtime Diagnostics before reporting a bug — it reports your device, backend, and which optional packages are installed.

Audio

MiniMax Music reads MINIMAX_API_KEY only from the ComfyUI server environment. It uses fixed global_en and cn_zh endpoints, supports music generation and cover models, decodes URL or hexadecimal responses, and emits MP3, WAV, or PCM results through the existing waveform and AUDIO sockets. The aigc_watermark field is sent only for cn_zh requests. See the official global or China music API reference for account and content requirements.

Legacy model loaders

Kept for existing workflows. New graphs should prefer Modern VLM, which covers most of these architectures through one interface.

Install

Install through ComfyUI Manager, or clone into ComfyUI/custom_nodes and run:

python -m pip install -r ComfyUI/custom_nodes/ComfyUI_VLM_nodes/requirements.txt

Run that command with ComfyUI's Python. Do not install or replace torch from this repository: ComfyUI's own installer selects CUDA, ROCm, XPU, Metal, or CPU. Current official bitsandbytes wheels are installed automatically only on their supported OS/architecture combinations. Unsupported machines retain all non-quantized nodes.

Robotics / VLA isolated runtimes

The robotics nodes keep policy dependencies outside ComfyUI. The universal HTTP client works without another package. Native openpi WebSocket and GR00T ZeroMQ clients use the lightweight optional extra:

python -m pip install \
  -r ComfyUI/custom_nodes/ComfyUI_VLM_nodes/requirements-robotics-client.txt

VLA Model Catalog covers current SmolVLA, X-VLA, π0/π0-FAST/π0.5, GR00T N1.7, WALL-OSS, MolmoAct2, VLA-JEPA, LingBot-VA, FastWAM, EO-1, EVO-1, OpenVLA-OFT, and Octo routes. “Available” means a supported isolated runtime/checkpoint path; base and architecture-only entries still require embodiment-specific training and transforms.

Start with SmolVLA for small consumer hardware. The included authenticated LeRobot sidecar loads one chosen policy, uses its serialized processors, returns action chunks over bounded JSON/JPEG, keeps it resident for speed, and can offload it to CPU after an idle timeout. Remote policy URLs require encrypted transport and explicit opt-in. Tokens are fixed environment variables (VLA_POLICY_TOKEN, OPENPI_API_KEY, or GROOT_API_TOKEN) and are never workflow inputs.

See examples/robotics/README.md for D-drive WSL setup, platform boundaries, current model readiness, observation schemas, action safety semantics, and the runnable API example.

Moondream 3 / 3.1 isolated runtime

Moondream's official Photon package pins Pillow below version 11 while current ComfyUI uses a newer Pillow. It therefore runs in a dedicated sidecar environment and never changes ComfyUI's Python packages. Read and accept the Moondream Model License 1.0, then create the environment under the registered LLavacheckpoints model folder.

Linux/WSL/macOS:

runtime="ComfyUI/models/LLavacheckpoints/moondream31-runtime"
uv venv "$runtime/.venv" --python 3.12
uv pip install --python "$runtime/.venv/bin/python" \
  -r ComfyUI/custom_nodes/ComfyUI_VLM_nodes/requirements-moondream31.txt

Windows PowerShell:

$runtime = "ComfyUI\models\LLavacheckpoints\moondream31-runtime"
uv venv "$runtime\.venv" --python 3.12
uv pip install --python "$runtime\.venv\Scripts\python.exe" `
  -r "ComfyUI\custom_nodes\ComfyUI_VLM_nodes\requirements-moondream31.txt"

The first Loader execution downloads the selected official model below that runtime's cache directory. Use moondream3.1-9B-A2B for query, caption, detection, and pointing. Use moondream3-preview only for the SVG segment skill; the final 3.1 model card does not list segment. Set the server-side MOONDREAM_PYTHON environment variable when using a different isolated environment. Do not put this path or any credential in a workflow.

Official Photon local inference currently supports NVIDIA Ampere-or-newer on Linux/Windows and Apple Silicon on macOS 13 or newer. It does not currently provide local ROCm, Intel GPU, or CPU execution. Those platforms retain every portable Transformers, GGUF, API, and vision utility node in this pack.

On CUDA 12 x86-64 systems the isolated requirements deliberately install nvidia-cuda-runtime-cu12==12.9.79. Kestrel 0.4.6's AOT kernels require the cudaLibraryLoadData entry point, which is absent from the CUDA 12.6 runtime bundled by cu126 PyTorch. This pin updates only Photon's private runtime; it does not replace ComfyUI's PyTorch build or the host NVIDIA driver.

GGUF nodes use optional llama-cpp-python. Install a wheel built for the desired CUDA, ROCm/HIP, Metal, Vulkan, SYCL, or CPU backend:

python -m pip install -r ComfyUI/custom_nodes/ComfyUI_VLM_nodes/requirements-llama-cpp.txt

See COMPATIBILITY.md for the tested matrix and official backend-specific GGUF commands.

The GGUF loaders now query the installed llama.cpp build instead of inferring its capabilities from PyTorch. Accelerator offload automatically falls back to CPU when a CPU-only wheel is installed. Advanced optional inputs expose logical and physical prompt batching (n_batch/n_ubatch), flash-attention policy, mmap, and CUDA/ROCm multi-GPU layer/row splitting without changing legacy workflow sockets. Auto flash attention retries the portable path if a backend/model pair rejects it.

The LLaVA Vision Projector Loader supports metadata-driven MTMD plus explicit handlers for LLaVA 1.5/1.6, MiniCPM-V 2.6, Moondream2, NanoLLaVA, Qwen2.5-VL, Gemma 4, Llama 3 Vision Alpha, and Obsidian. Use the default metadata-driven handler for current GGUF + mmproj pairs; select the named legacy handler when a model card requires it.

Models are downloaded only when their node first executes and are stored below ComfyUI/models/LLavacheckpoints. Hugging Face downloads respect HF_TOKEN. Gemma 3 and PaLI-Gemma require accepting their model licenses on Hugging Face.

GPU lifecycle

• ComfyUI managed (BF16) is the default and preferred path. BF16 is used only when the active device reports support; otherwise the node safely falls back to FP16 on CUDA/ROCm/Metal/XPU or FP32 on CPU.
• 4-bit/8-bit models and llama.cpp own external allocators. Before loading, the nodes ask ComfyUI to free the required space; unloading closes the exact owned model and then requests a soft cache cleanup. Small quantized models stay on ComfyUI's active device instead of assuming GPU zero. Large-model Accelerate placement is enabled on CUDA/ROCm/XPU; any disk offload remains inside the model's ComfyUI directory.
• llama.cpp model and projector bytes are included in the pre-load reservation. The runtime reports llama.cpp's own compiled backend, GPU-offload, mmap, and mlock capabilities in VLM Runtime Diagnostics.
• unload_after=false caches one model per node instance for fast repeated queues. Cache creation is serialized, so concurrent API work cannot make the same node allocate duplicate model handles. Turn it on for maximum reclamation between prompts.
• Moondream Photon asks ComfyUI to make room before it starts, then owns one exact isolated process. unload_after=true gracefully shuts it down and terminates that process if necessary, which releases Photon model, KV-cache, and CUDA-graph allocations without flushing unrelated ComfyUI models. The sidecar intentionally does not inherit ComfyUI's PyTorch allocator override; Photon's CUDA-graph capture uses the native allocator in its own process. The worker does not inherit unrelated provider keys or proxy credentials; only HF_TOKEN, and MOONDREAM_API_KEY for an explicitly selected adapter, may cross into its server-side environment. Base-model sidecars honor DO_NOT_TRACK locally and do not start Kestrel's anonymous telemetry task. Its random IPC secret is not placed on the process command line.
• A connected video_frames batch becomes the primary visual input. The optional still-image socket is ignored for video inference so smaller models cannot silently answer from the wrong media.
• Qwen 3.5/3.6 thinking is off by default for lower latency and predictable output length; enable it explicitly for tasks that benefit from visual reasoning.
• Auto (SDPA) is portable and preferred. Flash Attention 2 is accepted only on supported CUDA/ROCm builds and otherwise fails before model loading.
• VLM Runtime Diagnostics produces a zero-download JSON report containing OS, Python, PyTorch, backend, dtype capability, and optional package versions.
• Visualization-only companion repositories do not allocate accelerator memory.

Avoid placing several independently quantized VLMs in one workflow unless the GPU can hold them. On a 24 GB card, Qwen 3 VL 2B is the fast default, Qwen 3 VL 8B fits in BF16, and larger models should use NF4. Qwen 3.5/3.6 can be substantially slower when their optional optimized linear-attention kernels are not available for the installed PyTorch/backend combination.

API nodes

Hosted LLM API (Secure) and Hosted VLM API (Secure) share a provider layer built around the current OpenAI Responses and Chat Completions request shapes, with Anthropic using its native Messages/vision contract and Gemini switching to its native multimodal contract for grounded or structured calls. The VLM node accepts a still image or a video-frame batch, samples frames uniformly, resizes and JPEG-compresses them, and enforces per-image and total request limits before upload. Both nodes can stream text into a connected ViewText node.

Both API nodes also expose:

• Native web search for OpenAI, Gemini, Anthropic, xAI, and any compatible model routed through OpenRouter. Unsupported presets fail clearly before a model request instead of silently pretending to search. Search can add provider cost and has provider-specific data terms, so it is off by default.
• JSON object and JSON Schema output. Completed JSON is always parsed locally, JSON Schema results are validated locally, and invalid results fail the node instead of flowing into downstream automation.
• Open-source structured VLM output through Custom / Local endpoints. OpenAI-standard mode supports vLLM, Ollama, and compatible servers; llama.cpp JSON Schema emits llama.cpp's direct schema dialect; and JSON object + local validation is a portable fallback for servers that implement only JSON mode.

User-provided schemas are capped at 64,000 characters, bounded by depth/node count, checked against their declared JSON Schema draft, and may use only local fragment $ref values. Remote/file references are rejected so validation can never turn into an unexpected network or filesystem lookup.

Curated production profiles include:

Preset IDs were reviewed on 2026-07-29 against the official OpenAI, Gemini, Claude, xAI, DeepSeek, Groq, Mistral, and Together, plus OpenRouter's multimodal compatibility catalogs. Use model_override when a provider exposes a newer compatible model before the next node-pack release.

The capability routing follows the current official OpenAI web-search and structured-output contracts, Gemini grounding and structured output, Claude web-search and structured-output contracts, xAI web search and structured outputs, and OpenRouter server-side search. The local dialect is based on the llama.cpp server API.

API keys are not node inputs. A workflow contains only the provider selection, and the server resolves that provider's fixed environment variable at execution time. Built-in credentials are pinned to the provider's official HTTPS host; only the custom profile accepts a URL, and it can read only CUSTOM_API_KEY. Remote custom URLs require HTTPS, while keyless HTTP is restricted to localhost/loopback. Redirect following and environment proxies are disabled by default, API calls are stateless, OpenAI Responses explicitly use store=false, and provider exceptions are redacted before ComfyUI receives them.

Web search sends the prompt (and, where supported, the same multimodal request) to the selected provider's server-side search system. Do not enable it for content that must not be processed under that provider's search terms.

Opening an older PromptGenerateAPI workflow automatically clears its former plaintext key widget before the graph is configured. Save the migrated workflow to overwrite the old file, and rotate any key that was previously saved or shared. See SECURITY.md for setup and the exact threat model.

Reliability guarantees

• Importing the pack performs no network access, compilation, or package install.
• Missing optional backends fail only the node that needs them, with an actionable error.
• Image inputs use ComfyUI BHWC batches; text responses preserve every batch item. Florence/PaLI masks use BHW.
• forceInput string hacks were removed, preventing frontend widget-index drift.
• Downloads stay inside the configured ComfyUI model directory.
• CI installs and imports the full pack on Linux Python 3.10/3.13, Windows Python 3.12, and macOS Python 3.12. Backend contracts for CUDA, ROCm, Metal, XPU, and CPU are exercised without pretending hosted CPU runners are GPUs.

Run local checks with:

PYTHONPATH=/path/to:/path/to/ComfyUI python -m pytest -q

Real-weight checks are opt-in because they download multi-gigabyte checkpoints:

python tests/manual_model_smoke.py --model "Qwen 3 VL 4B Instruct"
python tests/manual_specialized_smoke.py --backend florence-large
python tests/manual_llama_cpp_smoke.py --download

See MODEL_VALIDATION.md for the exact real-weight and catalog-only evidence matrix.

Please report reproducible bugs at the issue tracker.

@software{Aydogan_ComfyUI_VLM_Nodes_2026,
  author  = {Aydoğan, Gökay},
  title   = {ComfyUI VLM Nodes},
  version = {3.5.0},
  year    = {2026},
  url     = {https://github.com/gokayfem/ComfyUI_VLM_nodes}
}